Multi-step time sequence prediction method and system based on double-segmentation conformal prediction

Abstract

The invention discloses a multi-step time sequence prediction method and system based on double-segmentation conformal prediction. The method comprises the following steps: acquiring new input data; distributing the new input data to a corresponding clustering cluster to obtain a clustering result; extracting information from the recorder according to a clustering result to construct a prediction interval; adjusting the corresponding content according to the prediction interval; wherein a similar trend sequence is vertically classified and clustered, errors of adjacent time steps are horizontally and dynamically combined to optimize window division, over-estimation and under-estimation errors are asymmetrically processed to construct a precise confidence interval, the precise confidence interval is stored in the recorder, and an error set is dynamically updated. By implementing the method provided by the invention, the defects in the prior art can be overcome through a two-dimensional segmentation mechanism, more accurate uncertainty quantization is realized, and the adaptability and accuracy of multi-step time sequence prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to a computer, more specifically to a multi-step time series prediction method and system based on double partition conformal prediction. BACKGROUND

[0002] Time series prediction has important application value in key fields such as energy system optimization and industrial process control, especially multi-step prediction can provide continuous future trend prediction for long-term decision-making. Uncertainty quantification (UQ) as a core technology to improve prediction reliability plays a crucial role in these fields. Conformal prediction (CP) has become a popular choice in the UQ field due to its independence on data distribution assumption and statistical guarantee.

[0003] However, the traditional CP framework and its improved version face three progressive technical bottlenecks in multi-step prediction scenarios: the mainstream CP variants (such as dynamic error update of EnbPI and α parameter adjustment of ACI) are essentially designed for single-step prediction. When applied to multi-step prediction, the recursive calling mechanism causes the mixing of error terms (ξ) between different prediction steps, causing confusion of error distribution. For example, in wind-solar power generation prediction, the mixing of error terms caused by changes in light intensity during dawn and dusk causes distortion of the confidence interval boundary. Horizontal partitioning methods (such as CopulaCPTS and CF-RNN) use static time window division methods, which are difficult to capture the distribution consistency of non-uniform periodic data. For example, in data center IT load prediction and solar energy supply prediction, the load fluctuation patterns of weekdays and holidays show asymmetric periodic characteristics, and the fragmented error set generated by fixed-length windows (such as 24-hour division) reduces statistical significance; while solar energy supply has a relatively stable periodicity (with 24 hours as a cycle), a more flexible time window adjustment strategy is needed. Existing methods fail to establish a scenario-aware error isolation mechanism, resulting in coupling of error terms in steady-state and transient processes. For example, the error distribution of three implicit working conditions of sunny, cloudy and rain in solar data set presents obvious multi-modal characteristics. The global error pooling strategy of traditional CP causes errors in different working conditions to contaminate each other, resulting in interval coverage failure in specific scenarios.

[0004] In actual industrial scenarios such as wind-solar power grid scheduling and data center PUE optimization, the above defects make the control strategy based on prediction face a double paradox: overly conservative confidence intervals lead to redundant energy storage system capacity, while overly tight intervals may cause voltage out-of-limit risk. This core contradiction is particularly prominent in non-stationary time series such as sudden load fluctuations and intermittent new energy output scenarios, severely restricting the effective transmission of prediction reliability to control actions.

[0005] To solve these problems, the academia attempts to improve from two directions: in the longitudinal dimension, CopulaCPTS models time step dependence through a copula function, but is limited by the Gaussian assumption; in the horizontal dimension, HopCPT introduces a modern Hopfield network to realize time sequence feature matching, but fails to solve the error distribution drift problem in multi-step prediction. Despite the progress, these improvements still do not build an error decoupling mechanism for multi-step prediction, and there are theoretical limitations in dynamic scene adaptability.

[0006] Therefore, it is necessary to design a new method to overcome the deficiencies in the prior art through a two-dimensional segmentation mechanism to realize more accurate uncertainty quantification and improve the adaptability and accuracy of multi-step time series prediction. SUMMARY

[0007] The purpose of the present application is to overcome the deficiencies of the prior art and provide a multi-step time series prediction method and system based on double segmentation conformal prediction.

[0008] To achieve the above-mentioned purpose, the following technical solutions are adopted: a multi-step time series prediction method based on double segmentation conformal prediction, comprising: obtaining new input data; allocating the new input data to corresponding clustering clusters to obtain clustering results; extracting information from the recorder according to the clustering results to construct a prediction interval; adjusting the corresponding content according to the prediction interval; wherein, by vertically classifying and clustering similar trend sequences, horizontally dynamically merging errors of adjacent time steps to optimize window division, asymmetrically processing overestimated and underestimated errors to construct accurate confidence intervals, storing in the recorder, and dynamically updating the error set.

[0009] The further technical solutions are: the accurate confidence interval is constructed by vertically classifying and clustering similar trend sequences, horizontally dynamically merging errors of adjacent time steps to optimize window division, asymmetrically processing overestimated and underestimated errors, storing in the recorder, and dynamically updating the error set, comprising: generating a plurality of clusters from historical prediction sequences through a clustering method, wherein each cluster contains sequences with similar trends and amplitudes; defining the similarity of errors of adjacent time steps in the same cluster for each time step error item in each cluster, completing dynamic window division for each cluster, wherein the error distribution in each window satisfies consistency; defining error items that meet corresponding set conditions as overestimated errors and underestimated errors, calculating the empirical quantiles of the upper and lower limits of the confidence interval, and constructing the confidence interval based on the empirical quantiles of the upper and lower limits of the confidence interval; storing the clusters, corresponding dynamic windows and confidence intervals in the recorder; updating error information in the recorder.

[0010] Further technical solutions are as follows: the historical prediction sequences are clustered into several clusters by a clustering method, including: cluster analysis is performed on the historical prediction sequences by a k-means algorithm and a silhouette coefficient evaluation to determine an optimal cluster number and assign a label; a similarity between a new prediction result and the historical prediction sequences is calculated by a soft dynamic time warping, samples meeting a requirement in similarity are selected, and a most frequent cluster to which the samples belong is assigned to the new prediction result; error items of the same cluster as the new prediction result are collected, and an asymmetric prediction interval is constructed according to the error items.

[0011] Further technical solutions are as follows: the error items of the same cluster as the new prediction result are collected, and an asymmetric prediction interval is constructed according to the error items, including: error items of all samples in each time step in the cluster to which the new prediction result belongs are collected to form an error subset; an asymmetric prediction interval is constructed for the new prediction result based on error distribution in the error subset.

[0012] Further technical solutions are as follows: for each error item in each time step in each cluster, a similarity between errors of adjacent time steps in the same cluster is defined, and dynamic window division is completed for each cluster, including: for each error item in each cluster, error subsets of adjacent time steps are compared in time sequence, a statistical difference between the error subsets of the adjacent time steps is calculated by a Kolmogorov-Smirnov test, and a p value is generated; starting from a first time step and traversing to a second last time step, whether adjacent time steps are merged into a window is determined according to the p value; if the p value exceeds a preset threshold, error items of the current and next time steps are merged; if the p value does not exceed the preset threshold, error items of an existing window are saved, and a new window is started from the current time as a starting point, and when the sequence end is reached, the last window is forced to be saved; for each window, error items of all time steps are integrated to form a merged error set, and a confidence interval of a corresponding time step of the window is constructed by using the merged error set.

[0013] Further technical solutions are as follows: the error items meeting corresponding set conditions are defined as overestimation errors and underestimation errors, empirical quantiles of upper and lower limits of the confidence interval are calculated respectively, and the confidence interval is constructed based on the empirical quantiles of the upper and lower limits of the confidence interval, including: The error term for each time step is directionally separated, and the error term is divided into overestimation error and underestimation error; Empirical quantiles of the overestimation error set and the underestimation error set are calculated to obtain empirical quantiles of upper and lower limits of the confidence interval; An asymmetric confidence interval is constructed based on the empirical quantiles of the upper and lower limits of the confidence interval.

[0014] Further technical solutions thereof are that the error information in the recorder is updated, and the error information comprises: An error term between the new prediction result and the true value is calculated, and the error term is added to the error subset of the corresponding cluster and window according to the vertical classification rule; When the size of the error subset in the window exceeds the preset maximum capacity, the oldest error term is automatically removed.

[0015] Further technical solutions thereof are that the new input data comprises historical weather data and power generation, and the prediction interval comprises a prediction interval of multi-step time series data.

[0016] Further technical solutions thereof are that the new input data comprises server load, environmental temperature and cooling system data, and the prediction interval comprises a prediction interval of multi-step load and temperature.

[0017] The application further provides a multi-step time series prediction system based on double-split conformal prediction, comprising: An acquisition unit is configured to acquire new input data; An allocation unit is configured to allocate the new input data to corresponding clustering clusters to obtain a clustering result; An extraction unit is configured to extract information from the recorder according to the clustering result to construct a prediction interval; An adjustment unit is configured to adjust corresponding content according to the prediction interval; Wherein, similar trend sequences are clustered by vertical classification, errors of adjacent time steps are dynamically merged horizontally to optimize window division, overestimation and underestimation errors are asymmetrically processed to construct an accurate confidence interval, the confidence interval is stored in the recorder, and the error set is dynamically updated.

[0018] The advantages of this invention compared to existing technologies are as follows: By introducing a dual-dimensional segmentation mechanism—vertical classification and horizontal dynamic merging—this invention effectively overcomes the shortcomings of existing technologies in multi-step time series prediction. Specifically, vertical classification assigns sequences to different clusters based on the trend similarity of the data, thereby capturing pattern changes in different scenarios; while horizontal dynamic merging compares and merges error distributions between adjacent time steps, optimizing window partitioning to improve prediction accuracy. Furthermore, by asymmetrically handling overestimation and underestimation errors, a more accurate confidence interval is constructed, enabling the model to not only adapt to different types of data features but also quantify the uncertainty in prediction. Based on this, the method achieves more accurate and reliable multi-step time series prediction, enhances the model's adaptability and prediction accuracy for various application scenarios, and continuously optimizes prediction performance through dynamically updating the error set. This method is particularly suitable for fields requiring high-precision prediction and stability, such as renewable energy generation prediction and data center energy management.

[0019] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the multi-step time series prediction method based on dual-segment conformal prediction provided in an embodiment of the present invention. Figure 2 A schematic diagram of the conformal prediction method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the overall process of the multi-step time series prediction method based on dual-segment conformal prediction provided in an embodiment of the present invention; Figure 4 for Figure 3 A flowchart detailing the calibration phase implementation; Figure 5 This is a schematic block diagram of a multi-step time series prediction system based on dual-segment conformal prediction provided in an embodiment of the present invention; Figure 6 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0022] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described, obviously, the described embodiments are a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work belong to the protection scope of the present application.

[0023] It should be understood that the terms "comprising" and "including" as used in the specification and the appended claims indicate the presence of the described features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0024] It should also be understood that the terms used in the present application specification are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the present application specification and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0025] It should be further understood that the term "and / or" as used in the present application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations thereof, and includes these combinations.

[0026] Please refer to Figure 1 , Figure 1 The flowchart of the multi-step time series prediction method based on double segmentation conformal prediction provided by the embodiments of the present application. The multi-step time series prediction method based on double segmentation conformal prediction is applied in a server, through a double segmentation mechanism, i.e. vertical classification clustering similar trend sequences and horizontal dynamic merging adjacent time step errors to optimize window division, combining asymmetric processing overestimation and underestimation errors to construct accurate confidence interval, to overcome the problems of inaccurate uncertainty quantification and poor adaptability in multi-step time series prediction in the prior art. Specifically, it first uses a clustering algorithm to group historical prediction sequences according to trend and amplitude similarity, and on this basis, dynamically adjusts the window size according to error similarity, to ensure that the errors in each window are consistent; then for each window, the empirical quantile of overestimation and underestimation errors is calculated respectively, so as to construct an asymmetric confidence interval, improving the accuracy and adaptability of the prediction interval. In addition, by updating the error set in real time and removing old error items, the latest and relevance of the model are maintained, further enhancing the accuracy of multi-step time series prediction.

[0027] Figure 1 The flowchart of the multi-step time series prediction method based on double segmentation conformal prediction provided by the embodiments of the present application. As Figure 1As shown, the method comprises the following steps S110-S140.

[0028] S110, obtaining new input data.

[0029] This step involves collecting new data for prediction. These data can be various types of time series data, such as historical weather data, power generation data, server load, environmental temperature and cooling system data, etc. These data are usually sourced from sensor networks, historical record databases or other real-time monitoring systems. The data obtained not only include observation values at the current time point, but also may contain continuous observation values in the past period of time, so as to provide sufficient context information for the model to make accurate prediction.

[0030] S120, assigning the new input data to corresponding clustering clusters to obtain clustering results.

[0031] In this stage, the vertical classification mechanism is used to assign the new input data to the most suitable clustering cluster. First, the optimal cluster number is determined by using the k-means algorithm combined with the silhouette coefficient evaluation, and the historical prediction sequence is clustered. Then, the similarity between the new prediction result and the historical prediction sequence is calculated by using soft dynamic time warping (Soft-DTW), and the samples with required similarity are selected and assigned to the most frequently occurring cluster label. This step ensures that the new data can be reasonably classified into the error distribution group with similar trend and amplitude, thereby improving the accuracy of subsequent prediction.

[0032] S130, constructing a prediction interval from the recorder according to the clustering results.

[0033] Based on the clustering results obtained in step S120, the system retrieves the corresponding error set from the recorder to construct the prediction interval for the new input data. Specifically, for each time step within a cluster, the Kolmogorov-Smirnov test is used to dynamically divide the error window, and the empirical quantiles of overestimation error and underestimation error are calculated according to the error distribution within these windows. Subsequently, these empirical quantiles are used to construct an asymmetric confidence interval, which takes into account the possibility of the predicted value being higher than the true value, as well as the possibility of being lower than the true value. This processing method helps to generate a more compact and higher coverage probability prediction interval.

[0034] The new input data includes historical weather data, power generation, and the prediction interval includes prediction intervals of multi-step time series data.

[0035] The new input data includes server load, environmental temperature and cooling system data, and the prediction interval includes prediction intervals of multi-step load and temperature.

[0036] S140, adjust the corresponding content according to the prediction interval.

[0037] Finally, according to the generated prediction interval, the user can make corresponding decisions or adjustment strategies according to specific business scenarios. For example, in power grid dispatching, the "worst-case value" of the power generation prediction interval can be used to reserve standby capacity, and the "best-case value" can be used to optimize the energy storage charging and discharging strategy; in data center management, the cooling power can be dynamically adjusted according to the temperature prediction interval to avoid overcooling or equipment overheating and other problems. In addition, when the prediction interval suddenly expands, a device inspection process can be automatically triggered to prevent potential abnormal situations. In this way, the uncertainty in multi-step time series prediction is precisely managed and controlled, improving resource utilization efficiency and system stability.

[0038] In summary, the above four steps together constitute a complete workflow of the double-split conformal prediction (DSCP) method. By innovatively introducing a vertical-horizontal dual-dimension error decoupling mechanism, the error distribution confusion problem existing in the prior art is effectively solved, and the accuracy and adaptability of multi-step prediction are significantly improved.

[0039] Among them, by vertically classifying and clustering similar trend sequences, horizontally dynamically merging errors of adjacent time steps to optimize window division, asymmetrically processing overestimation and underestimation errors to construct accurate confidence intervals, storing them in the recorder, and dynamically updating the error set.

[0040] In an embodiment, the above-mentioned construction of accurate confidence intervals by vertically classifying and clustering similar trend sequences, horizontally dynamically merging errors of adjacent time steps to optimize window division, asymmetrically processing overestimation and underestimation errors, storing them in the recorder, and dynamically updating the error set includes steps S151-S155.

[0041] S151, generate a number of clusters for historical prediction sequences by clustering method, wherein each cluster contains sequences with similar trends and amplitudes.

[0042] For historical prediction sequences , generate clusters by k-means clustering method, denoted as Each cluster contains sequences with similar trends and amplitudes.

[0043] In an embodiment, the above-mentioned step S151 can include steps S1511-S1513.

[0044] S1511, perform clustering analysis on historical prediction sequences by k-means algorithm and silhouette coefficient evaluation to determine the optimal number of clusters and assign labels.

[0045] In this embodiment, in this step, firstly, the historical prediction sequences are preliminarily clustered using the k-means algorithm. In order to ensure the quality of the clustering results, the silhouette score is used as an evaluation criterion to determine the optimal cluster number. The specific operation is as follows: For different cluster numbers k (from 1 to the maximum preset value N), the k-means clustering is performed respectively.

[0046] The silhouette score of each clustering result is calculated, and the k value that makes the silhouette score maximum is selected as the optimal cluster number.

[0047] Finally, the historical prediction sequences are divided into k independent error categories, and each sample is assigned a corresponding cluster label.

[0048] S1512, the similarity between the new prediction result and the historical prediction sequence is calculated by using soft dynamic time warping, samples meeting the requirement are selected, and the most frequent cluster to which the sample belongs is assigned to the new prediction result.

[0049] In this embodiment, the new prediction result refers to a prediction value sequence of a future time point or time period.

[0050] In this step, the main purpose is to assign the new prediction result to a suitable cluster. The specific process is as follows: The similarity scores between the new prediction result and all historical prediction sequences are calculated using the soft dynamic time warping (Soft-DTW) algorithm.

[0051] The similarity scores are arranged in descending order, and the historical prediction samples corresponding to the top s highest scores are selected, wherein s is the set minimum cluster sample quantity.

[0052] The most frequent cluster label in these high-similarity samples is counted, and is assigned to the new prediction result.

[0053] S1513, collect the error items of the same cluster as the new prediction result, and construct an asymmetric prediction interval according to the error items.

[0054] In this embodiment, the asymmetric prediction interval refers to a confidence interval that contains upper and lower bounds and considers different risks of overestimation and underestimation for each prediction value according to the error distribution.

[0055] In an embodiment, the above step S1513 can include steps S15131-S15132.

[0056] S15131, for the cluster to which the new prediction result belongs, collect the error items of all samples in the cluster at each time step to form an error subset.

[0057] In this embodiment, the error subset refers to a set of prediction errors at all time steps associated with the samples within a specific cluster extracted from historical data, used to construct an asymmetric prediction interval.

[0058] The task of this stage is to extract relevant information from historical data for subsequent construction of prediction intervals. The specific operations include: Determine the cluster to which the new prediction result belongs.

[0059] Extract the error terms of all samples within the cluster at each time step to form an error subset. This error subset contains all error information within the cluster for subsequent analysis.

[0060] S15132, based on the error distribution in the error subset, construct an asymmetric prediction interval for the new prediction result.

[0061] In this step, the error subset collected is used to construct an asymmetric prediction interval, and the specific process is as follows: Separate the direction of error terms and define overestimation error (when the true value is lower than the predicted value) and underestimation error (when the true value is higher than the predicted value).

[0062] Calculate the empirical quantile of overestimation error and underestimation error, i.e. the upper and lower bounds of the confidence interval.

[0063] Construct an asymmetric confidence interval.

[0064] Through the above steps, the DSCP method can effectively quantify and manage the uncertainty in multi-step time series prediction. It not only solves the confusion problem of error distribution in traditional conformal prediction, but also provides more accurate and adaptive prediction intervals. This method is particularly suitable for application scenarios that require fine management and control of uncertainty, such as renewable energy generation prediction, data center energy management, etc. In addition, its model independence and flexible integration make it widely applicable to various industries, significantly improving the engineering applicability and practical value of multi-step prediction.

[0065] In this embodiment, first, clustering initialization is performed: for historical prediction sequences , use the k-means algorithm for clustering, test the number of clusters from 1 to the maximum preset value , generate cluster results.

[0066] Then perform quality assessment: calculate the silhouette coefficient of each clustering result, which is defined as: ; where is the average distance of sample from other samples in the same cluster when the number of clustering clusters is , For the average distance from the cluster center to the nearest outlier sample. Choose the value of as the optimal number of clusters, and output the final clustering labels , where represents the label of the cluster to which the sample belongs. For the new prediction , calculate its similarity score with all historical predictions using Soft-DTW, where . Then perform nearest neighbor screening, sort in descending order, and select the top historical predictions corresponding to the largest scores, where is the sample size of the smallest cluster. Let the set of screened sample indices be . Subsequently, perform cluster assignment by counting the cluster label frequencies of the samples in , and assign to the cluster with the highest frequency, denoted as , i.e., ; where is the indicator function. Extract all sample index sets = from historical predictions that are in the same cluster as , and collect their corresponding error terms to construct the error subset: ; where is the prediction window length, denotes the error of the th historical prediction at the th time step. Subsequently, construct intervals using the error distribution in

[0067] to build asymmetric prediction intervals for .

[0068] Thus, the problem of cross-scenario prediction error contamination, such as multi-cloud weather error interfering with clear day predictions, is solved.

[0069] S152、For each cluster, define the error similarity between adjacent time steps in the same cluster for each time step, and complete dynamic window partitioning for each cluster, where the error distribution within each window satisfies consistency. ​​​​​

[0070] In one embodiment, step S152 described above may include steps S1521 to S1523.

[0071] S1521. For each error term within a cluster, compare the error subsets of adjacent time steps in chronological order, calculate the statistical difference between the error subsets of adjacent time steps using the Kolmogorov-Smirnov test, and generate a p-value.

[0072] In this embodiment, the p-value is a measure used by the Kolmogorov-Smirnov (KS) test to compare whether there are significant differences between error subsets at adjacent time steps. The smaller the p-value, the lower the probability that the two sets of data come from the same distribution; conversely, a higher p-value means that the two sets of data may have the same error distribution characteristics.

[0073] S1522. Starting from the first time step, traverse to the second to last time step. Determine whether to merge adjacent time steps into one window based on the p value. If the p value exceeds a preset threshold, merge the error terms of the current and next time steps. If the p value does not exceed the preset threshold, save the error terms of the existing window and open a new window starting from the current time. When the end of the sequence is reached, the last window is forcibly saved.

[0074] Specifically, for each error term within a cluster, all time steps are first traversed in chronological order, starting from the first time step and continuing to the second-to-last time step.

[0075] Perform a KS test between each pair of adjacent time steps to obtain their p-values.

[0076] If the p-value exceeds the preset threshold (usually a value between 0 and 1 used to control the requirements for merging similarity), then the error distributions of the two time steps are considered to be similar enough to be merged into one window.

[0077] If the p-value does not exceed the preset threshold, the existing window remains unchanged, and a new window is opened starting from the current time.

[0078] When the traversal reaches the end of the sequence, regardless of the last p value, it is necessary to forcibly save the error term set of the last window.

[0079] S1523. For each window, integrate the error terms of all time steps to form a merged error set, and construct the confidence interval of the time step corresponding to the window using the merged error set.

[0080] For each finalized window, the error terms of all time steps contained therein are integrated to form a merged error set.

[0081] The upper and lower limits of the asymmetric prediction interval are calculated using this merged error set, i.e., the empirical quantiles of the overestimation error and underestimation error, respectively.

[0082] Based on these quantiles, asymmetric confidence intervals are constructed for the time steps corresponding to the window, providing more accurate risk assessment and decision support.

[0083] Through the above process, the DSCP method can effectively identify and separate error patterns under different characteristics, thereby improving the accuracy and reliability of multi-step time series prediction. This method is particularly suitable for application scenarios that require fine-grained management of uncertainty, such as renewable energy generation prediction, data center energy management, and other fields.

[0084] In this embodiment, for each cluster , the error sequence is traversed in time step order adjacent to and , and the following operations are performed: First, K-S test calculation: for the error subsets and of time steps and , calculate the two-sample Kolmogorov-Smirnov statistic: ; where and are the empirical cumulative distribution functions of and , respectively, is the subset sample size.

[0085] Subsequently, the value generation: based on , calculate the p-value , quantifying the probability that and come from the same distribution.

[0086] Initialize the starting point of the current window , traverse time steps to , and merge the window according to the following rules: Merge condition judgment: if (the range is 0-1, used to control the similarity requirement of merging), then merge time steps and , and update the window coverage: ; Window segmentation operation: if , save the merged error set of all time steps in the current window ​ and reset the window start point as .

[0087] Termination: when the end time step is reached , the merging result of the last window is forced to be saved.

[0088] For each final window , its merging error set is defined as: ; i.e. all time steps within the window share the same merging error set . This set will be used to construct the confidence interval for the corresponding time step , in the form of: ; where is the -quantile function of the error set.

[0089] In this way, the prediction error information of e.g. “stable period in the early morning” and “mutant period in the morning rush hour” is not mixed in statistics, reducing the effect of uncertainty quantification.

[0090] S153, define the error items satisfying the corresponding set conditions as overestimation errors and underestimation errors, respectively calculate the empirical quantiles of the upper and lower limits of the confidence interval, and construct the confidence interval based on the empirical quantiles of the upper and lower limits of the confidence interval.

[0091] In this embodiment, the confidence interval refers to the range formed by the upper and lower limit empirical quantiles calculated respectively after separating the prediction error directionality (overestimation error and underestimation error) of each time step.

[0092] In an embodiment, the above step S153 can include steps S1531-S1533.

[0093] S1531, directionally separate the error items for each time step, and divide the error items into overestimation errors and underestimation errors.

[0094] This step first performs a sign analysis on the prediction error of each time step, and divides the error into two categories: when the actual value is less than the predicted value, it is defined as an overestimation error ; when the actual value is greater than the predicted value, it is defined as an underestimation error . This step ensures that errors of different directions are processed separately to more accurately reflect the uncertainty of the prediction.

[0095] S1532, calculate the empirical quantiles of the overestimation error set and the underestimation error set to obtain the empirical quantiles of the upper and lower limits of the confidence interval.

[0096] Calculate the empirical quantiles corresponding to the specified confidence level from both the overestimation and underestimation error sets. For the overestimation error set, calculate the coverage probability as follows: The upper bound quantiles are calculated for the underestimation error set; for the underestimation error set, the lower bound quantiles with a coverage probability of α / 2 are calculated. These quantiles represent the expected error boundaries at a given confidence level.

[0097] S1533. Construct asymmetric confidence intervals based on empirical quantiles of the upper and lower limits of the confidence interval.

[0098] The final step is to construct an asymmetric confidence interval using the upper and lower bound empirical quantiles obtained in the previous steps. Specifically, this involves adding or subtracting the corresponding quantile values ​​from the predicted point value to form an interval that reflects the potential range of fluctuations in the actual future value. The upper bound of this interval is composed of the predicted point value minus the quantile for underestimation error, while the lower bound is composed of the predicted point value plus the quantile for overestimation error. This confidence interval not only considers the uncertainty in the forecast but also pays special attention to different types of forecast errors, thus providing more accurate risk assessment and decision support.

[0099] In this way, the method effectively solves the interval expansion problem in traditional conformal prediction methods and improves the accuracy and dynamic adaptability of multi-step time series prediction.

[0100] Specifically, the prediction error at each time step is directionally separated, and the overestimation error is defined. with underestimation error for: ,in For predicted values, The true value. Through this decomposition, the original error set... Divided into two subsets: ; This step ensures that the direction information of the error is explicitly preserved.

[0101] right and Calculate their empirical quantiles respectively: where the upper bound quantile of the confidence interval is obtained by extracting the coverage probability from the overestimation error set. Quantile construction: ; The lower bound quantile is obtained by extracting the coverage probability from the underestimation error set. Quantiles: ;in This is an indicator function.

[0102] By overlaying quantile boundaries onto the predicted point values, asymmetric intervals are generated. ; This design satisfies the coverage probability constraint: ; Thus, the prediction error information of, for example, the "early morning stable period" and the "morning peak mutation period" is not mixed for statistics, and the uncertainty quantification effect is reduced.

[0103] S154, storing the cluster, the corresponding dynamic window, and the confidence interval in the recorder; S155, updating the error information in the recorder.

[0104] In an embodiment, the above-mentioned step S155 can include steps S151-S152.

[0105] S151, calculating an error term between the new prediction result and the true value, and adding the error term to the error subset of the corresponding cluster and window according to the vertical classification rule; S152, when the size of the error subset in the window exceeds the preset maximum capacity, automatically removing the oldest error term.

[0106] Specifically, when the new prediction result arrives, the true value of the new prediction result is compared with the true value of the previous prediction result: First, the current error term is generated; then the cluster and the corresponding window are determined according to the vertical classification rule in claim 2; then the is updated to the error set of the corresponding window, that is: ; The maximum capacity of each window is preset (for example ), when the size of the window error set exceeds the threshold: ; ; This operation ensures that the error set only retains the most recent samples, and reduces the negative impact of historical obsolete data on the current situation.

[0107] Thus, in the device aging scenario, the error distribution change is automatically tracked, and manual parameter adjustment is not required.

[0108] In this embodiment, the time series data is cleaned and standardized, and the training set, calibration set and test set are divided; k-means clustering and category assignment are performed; the time window is dynamically divided based on the KS test; the quantile is calculated and the prediction interval is output; the error set and the model parameters are updated in real time.

[0109] Among them, the training set (Training set): the data set used to train the point prediction model. ​

[0110] Calibration set: dataset used to compute error terms and build error set .

[0111] Testing set: dataset used to evaluate conformal prediction (CP) performance.

[0112] : multi-step time series input data, where denotes the dimensional feature vector at time step .

[0113] : target value sequence, denotes the real observation at time step .

[0114] : input window data, dimension , where is the prediction input window size of the point prediction model .

[0115] : point prediction model 's prediction result for input at steps.

[0116] : error term, defined as the difference between the real value and the predicted value , biased towards the error value at a certain time step of a certain prediction.

[0117] : error subset, the set of error terms computed from the calibration set, used to build prediction intervals.

[0118] Error pool refers to the merged set of error subsets.

[0119] The prediction error is biased towards one or more error values that are not specifically indicated.

[0120] The traditional method mostly uses the entire error set, while the present embodiment separates a subset from the error set for use.

[0121] : positive error (error term where predicted value is higher than real value). : negative error (error term where predicted value is lower than real value). : prediction interval with confidence level , whose probability of covering the real value is no less than .

[0122] : quantile of error set : quantile of error set

[0123] : quantile of error set : quantile of error set

[0124] : interval form: upper and lower bounds of prediction interval are calculated by subtracting quantiles of error set : interval form: upper and lower bounds of prediction interval are calculated by subtracting quantiles of error set

[0125] : time steps of input window

[0126] : step length of multi-step prediction

[0127] : input feature dimension

[0128] : user-defined significance level, controls the balance between coverage probability and width of prediction interval. (usually take 0.1-0.2, which means that the resulting confidence interval has a 90%-80% probability of covering the future true value) The method of the embodiment is realized through the two mechanisms of vertical classification and horizontal dynamic merging: first, based on the similarity of the trend and amplitude of historical prediction results, k-means clustering is used to classify the prediction vertically, ensuring consistent error distribution within the same category; then, through Kolmogorov-Smirnov test, the similarity of error distribution of adjacent time steps is dynamically evaluated, and time steps with similar errors are merged into the same time window to avoid error interference across time windows. In addition, DSCP introduces asymmetric error processing to distinguish between overestimation error and underestimation error , and constructs a prediction interval with balanced tightness and coverage. This method is model-independent, supports any prediction algorithm, and continuously optimizes through dynamic error set updating, solving the problems of insufficient statistical information and interval redundancy caused by mixed heterogeneous errors in traditional conformal prediction in multi-step scenarios.

[0129] Traditional prediction models (such as LSTM, XGBoost) only output a single prediction value, making it difficult to assess risk. The DSCP system generates statistically reliable confidence intervals for each step of prediction through dynamic error decoupling technology, directly empowering the following scenarios: ​​​​Risk Controllable Decision Making: For example, in power grid dispatch, reserve capacity is reserved based on the "worst-case" prediction of power generation, and the charging and discharging strategy of energy storage is optimized based on the "best-case".

[0130] Efficiency Improvement of Resource Utilization: For example, the cooling system of a data center dynamically adjusts the refrigeration power according to the temperature prediction interval, avoiding excessive refrigeration or equipment overheating.

[0131] Early Warning of Faults: For example, when the prediction interval suddenly expands, an automatic device inspection process is triggered to prevent potential abnormalities.

[0132] Core capabilities of the DSCP system in this embodiment: Scenario Adaptive Classification (Vertical Segmentation): Based on the dynamic time warping (DTW) metric to measure the morphological similarity of prediction sequences, combined with unsupervised clustering optimized by contour coefficients, historical predictions are divided into scenario categories with homogeneous error distribution (such as weather patterns, load conditions).

[0133] Error Isolation Reduces System Risk: The error distribution of independent scenarios meets the Homoscedasticity assumption, avoiding Covariate Shift across scenarios.

[0134] Reducing Model Re-training Frequency: The data distribution within the same scenario is more stable, and the model degradation rate is slowed down.

[0135] Supporting Non-stationary Scenario Expansion: New scenarios (such as typhoon patterns) can dynamically generate independent error clusters without the need for global reconstruction.

[0136] Dynamic Window Merging (Horizontal Segmentation): Based on the p-value threshold mechanism of Kolmogorov-Smirnov two-sample test, the consistency of error distribution between adjacent time steps is judged, and dynamic expansion and segmentation of windows are realized.

[0137] Balancing Statistical Power and Computational Efficiency: The sample size of the merged window meets the Central Limit Theorem (CLT), and the variance of quantile estimation is reduced.

[0138] Adaptive Data Flow Changes: KS test is sensitive to distribution shape differences and can capture implicit concept drift.

[0139] Reducing Storage Overhead: Merged windows share the same error distribution, eliminating the need to store historical data independently for each time step.

[0140] Asymmetric Error Handling: Distinguishing between "prediction high" and "prediction low" errors, generating independent confidence intervals for upper and lower bounds, breaking through the information bottleneck of traditional symmetric intervals. Enhancing the practicality of risk-sensitive scenarios: Asymmetric quantiles meet the needs of unilateral risk control (such as financial downside risk and grid under-generation penalties).

[0141] Mitigating the problem of excessive conservatism in intervals: The distribution skewness of overestimation and underestimation errors is fundamentally different, and symmetrical processing leads to interval redundancy.

[0142] Supports customized business strategies: By adjusting the upper and lower bounds of the α parameter, the tolerance for overestimation / underestimation risk can be independently controlled.

[0143] By encapsulating the predicted value interface, the output of the prediction model at any point is mapped to the input of the DSCP system, thus achieving algorithm independence.

[0144] Protect existing asset investments: Compatibility design follows the Open-Closed Principle, ensuring that scalability does not affect the original system.

[0145] Accelerate cross-domain migration: Abstract interfaces decouple business logic from the statistical computing layer.

[0146] Reduce technology stack dependencies: Standardize input / output formats to mask framework differences.

[0147] contrast Figure 2 as well as Figure 3 The DSCP method in this embodiment has a workflow divided into three stages: calibration, deployment, and update. During the calibration stage... - The builder module uses self-clustering units to process historical prediction calibration sets. Horizontal segmentation is performed to divide the data by category; then, within each category, the time step information is vertically segmented using a dynamic time step merging unit to dynamically allocate the errors of different time steps. After being categorized and dynamically merged by time step, the results are stored in [location missing]. - Logger. During the deployment phase, The classifier will take new input data Assign to the corresponding cluster. - The selector then selects from the clustering results... - Extract information from the recorder to construct the prediction interval. The update phase compares the predicted values... Compared with measured data Dynamic updates - Error information in the collector. This process achieves fine-grained management of cross-category feature and time dimension errors through a dual horizontal-vertical segmentation strategy.

[0148] This process achieves refined management of cross-category features and time dimension errors through a dual horizontal and vertical segmentation strategy.

[0149] Please see Figure 4 In the self-clustering stage, appropriate classification methods are used to integrate historical predictions. Divided into groups 1 to 2 Predictions within the same category are assigned the same color. In the dynamic merging phase, the KS test is used to evaluate the correlation between adjacent time steps within each category. Distribution similarity.

[0150] The method in this embodiment addresses the following technical problems: existing conformal prediction methods suffer from error distribution confusion and interval expansion due to strong real-time data dependence in multi-step time-series prediction scenarios; and they also lack the technical ability to balance prediction interval accuracy and dynamic adaptability in industrial applications. Specifically, to solve these problems, the method analyzes historical multi-step prediction results... (in Indicates the first Prediction windows Vertical feature clustering is performed on the predicted values, and a similarity matrix is ​​constructed using the Dynamic Time Warping (DTW) distance. (in The optimal number of clusters is determined based on the silhouette score criterion. (in To preset the maximum number of clusters, (for k-means clustering silhouette coefficients), ultimately... Divided into Individual error categories ( ), including new predictions Nearest neighbor search via soft-DTW ( ) Activate the category to which the historically similar sequence belongs .

[0151] Within each category, horizontal error distributions are merged, and a statistically consistent time series window is dynamically aggregated using the p-value threshold mechanism of the Kolmogorov-Smirnov test to generate an error subset with spatiotemporal correlation.

[0152] For the multi-step prediction sequence to be tested, the similarity features between it and the historical prediction are matched by soft dynamic time warping (Soft-DTW), and the dynamic merging error subset under the corresponding error category is activated. Asymmetric prediction intervals are constructed based on quantile regression, and calculations are performed independently for each time window. , Quantile boundaries enable the generation of error sign-sensitive intervals.

[0153] Furthermore, the system corresponding to this embodiment includes: A vertical classification engine module configured to perform a magnitude-trend joint clustering of the prediction sequence to generate error class labels with operating condition discrimination capability; A dynamic aggregation processor module with a K-S test unit and a sliding window controller to realize dynamic aggregation and update of error distribution; An interval generator module integrating an asymmetric quantile calculation unit and a real-time feedback calibration interface to output dynamically adaptive multi-step prediction intervals.

[0154] The method of the embodiment overcomes the core problem of error heterogeneous distribution interference in multi-step prediction by innovatively introducing a vertical-horizontal two-dimensional error decoupling mechanism. Compared with existing technologies, the following breakthrough effects are achieved: Prediction interval precision improvement: On solar and wind farm energy data sets, the Winkler score is improved by 23.59% compared to the optimal baseline method, and the interval width is reduced by 30.14%; Dynamic adaptability enhancement: Through K-S test driven error subset dynamic aggregation, stable prediction coverage is maintained in data distribution drift scenarios; Industrial-level energy efficiency optimization: Applied to the simulation of data center carbon footprint management scenarios, through accurate multi-step renewable energy output prediction intervals, the cooling system scheduling strategy is optimized to achieve an 11.25% carbon emission reduction optimization effect.

[0155] Significantly improves the engineering applicability of multi-step prediction in key fields such as smart grid and industrial Internet of Things, and provides statistical reliability guarantee for prediction-driven dynamic resource scheduling.

[0156] The method of the embodiment is based on the morphological similarity of historical prediction sequences (such as the DTW distance of time series data), and uses clustering methods (such as k-means) to divide the data into multiple scene categories (such as sunny / cloudy patterns, which may be divided based on numerical conditions rather than specific real weather conditions), to isolate error distribution.

[0157] Kolmogorov-Smirnov test is used to dynamically evaluate the similarity of error distribution of adjacent time steps, and consistent time windows are merged to balance statistical effectiveness and dynamic adaptability.

[0158] Asymmetric error processing: distinguishing overestimation errors and underestimation errors , independently calculate upper and lower bound quantiles to generate tighter prediction intervals, avoiding information loss of traditional symmetric intervals.

[0159] Dynamic update and model independence: Through incremental error injection and sliding window control, the error set is optimized in real time; supports any point prediction model (such as LSTM, XGBoost), reduces deployment dependence.

[0160] The application scenarios can cover specific application fields such as renewable energy power generation prediction and data center energy management.

[0161] The prior art defects such as the HopCPT scheme only perform horizontal segmentation, and do not isolate multi-scene errors (such as the mixing of prediction errors of sunny and cloudy weather), resulting in excessive conservatism or insufficient coverage of the interval.

[0162] The embodiment isolates multi-scene errors through vertical classification, and combines horizontal dynamic merging to expand the statistical sample size. The historical prediction is clustered according to the trend / magnitude (such as distinguishing weekday and holiday patterns), ensuring that each error distribution is homogeneous, and fundamentally solving the problem of covariate shift. The result is that the prediction interval width is reduced by an average of 30.14%, and the Winkler score is improved by 23.59%.

[0163] The prior art defects such as the CF-RNN method use fixed time steps and independent windows, ignoring the distribution consistency of adjacent time steps (such as forcibly dividing the stable period in the early morning and the sudden change period in the morning peak into several time windows), resulting in fragmentation of statistical samples.

[0164] The embodiment dynamically merges similar consecutive time steps through K-S test (such as merging the stable period throughout the day), balancing statistical effectiveness and dynamic response capability. For periodic data (such as wind speed, load fluctuations), DSCP maintains stable coverage, while existing methods have a sharp expansion of the interval when the cycle mutates.

[0165] The prior art defects such as traditional CP symmetric intervals cannot distinguish between overestimation and underestimation risks, resulting in interval redundancy or risk control failure.

[0166] The embodiment separates and and independently calculates the quantile to generate an interval with independent upper and lower bounds. In wind farm prediction, the upper bound of DSCP (to cope with the risk of insufficient power generation) is significantly narrower than the lower bound, which is more in line with actual business needs.

[0167] The prior art defects such as the mainstream method has large computational overhead; CQR requires retraining of quantile models, which destroys model independence.

[0168] The embodiment supports any point prediction model (such as LSTM, XGBoost, etc.), has flexible deployment capabilities; clustering and K-S test are both parameter-free statistical methods, with low average single-step processing time during deployment; uses an incremental error injection mechanism to capture the temporal changes of data streams.

[0169] On the basis of horizontal segmentation of HopCPT, artificial scene classification based on business rules is added (such as pre-grouping data according to weather types). The limitations are that it relies on expert knowledge and cannot adaptively discover potential scene patterns; static classification is difficult to cope with unknown scenes (such as new extreme weather); and there is a lack of distribution consistency test, and window division is still rigid.

[0170] Train multi-task models that share underlying features and have independent output heads to handle different scenarios. Challenges include the need for large amounts of labeled data to support multi-task training; performance drops sharply when scene boundaries are ambiguous; and lack of statistical guarantee theory support for DSCP.

[0171] Construct hierarchical priors to automatically identify scenes and estimate uncertainty. Challenges include complex inference processes and high computational overhead; strong subjectivity in prior settings, low engineering practicability; and difficulty in implementing real-time updating mechanisms for DSCP.

[0172] Combine multiple specialized models (such as sunny and rainy weather predictors) and dynamically weight them based on real-time features.

[0173] Meta-learning framework: quickly adapt to new scenarios using models, but still need to solve the problem of uncertainty quantification.

[0174] Attention mechanism + interpretable AI: identify key scene features through interpretable analysis to assist in optimizing artificial rules.

[0175] In the application of renewable energy generation prediction, this method first inputs historical weather data and power generation time series data, and uses a double segmentation mechanism (vertical classification and horizontal dynamic merging) to process these data, outputting prediction intervals for multiple time series data. Vertical classification divides data into different weather scene categories, such as sunny or cloudy patterns, based on the morphological similarity of historical data; while horizontal dynamic merging evaluates the error distribution similarity between adjacent time steps to merge consistent time windows, thereby generating more accurate prediction intervals. Next, based on these prediction intervals, the power grid dispatching strategy is optimized, through identifying the stability and volatility of power generation in different time periods, to achieve precise control of the power system, not only improving power generation stability, but also improving energy utilization.

[0176] For the application of data center energy management, the method focuses on inputting time series data such as server load, environmental temperature and cooling system operating parameters, and using the same double-split conformal prediction process to output the prediction interval of multi-step load and temperature change. This process can help managers better understand the thermal load situation and its trend in the data center in the future period. According to the prediction interval, the setting parameters of the cooling system, such as supply air temperature and mass flow rate, can be dynamically adjusted to adapt to the real-time changing thermal load demand, ensuring the cooling effect while reducing unnecessary energy consumption. In addition, this method helps to reduce the overall carbon emissions of the data center and improve energy efficiency, thereby achieving the goal of green data center. In this way, not only the safe and stable operation of the data center is ensured, but also the energy saving and emission reduction target is achieved.

[0177] The above multi-step time series prediction method based on double-split conformal prediction effectively overcomes the shortcomings of existing technologies in multi-step time series prediction by introducing a double-dimensional splitting mechanism, vertical classification and horizontal dynamic merging. Specifically, vertical classification allocates sequences to different clustering clusters according to the trend similarity of data, thereby capturing pattern changes in different scenarios; while horizontal dynamic merging compares and merges the error distribution between adjacent time steps, optimizing window division to improve prediction accuracy. In addition, by asymmetrically processing overestimated and underestimated errors, a more accurate confidence interval is constructed, so that the model can not only adapt to different types of data characteristics, but also quantify the uncertainty in prediction. Based on this, the method realizes more accurate and reliable multi-step time series prediction, enhances the adaptability and prediction accuracy of the model to various application scenarios, and continuously optimizes the prediction performance through dynamic updating of error sets. This method is particularly suitable for fields that require high-precision prediction and stability, such as renewable energy generation prediction and data center energy management.

[0178] Figure 5 is a schematic block diagram of a multi-step time series prediction system 300 based on double-split conformal prediction provided by an embodiment of the present application. As Figure 5 shown, corresponding to the above multi-step time series prediction method based on double-split conformal prediction, the present application also provides a multi-step time series prediction system 300 based on double-split conformal prediction. The multi-step time series prediction system 300 based on double-split conformal prediction includes units for performing the above multi-step time series prediction method based on double-split conformal prediction, and the system can be configured in a server. Specifically, please refer to Figure 5 , the multi-step time series prediction system 300 based on double-split conformal prediction includes an acquisition unit 301, an allocation unit 302, an extraction unit 303 and an adjustment unit 304.

[0179] The acquisition unit 301 is configured to acquire new input data; the allocation unit 302 is configured to allocate the new input data to a corresponding cluster to obtain a clustering result; the extraction unit 303 is configured to extract information from the recorder according to the clustering result to construct a prediction interval; and the adjustment unit 304 is configured to adjust the corresponding content according to the prediction interval. The vertical classification is used to cluster similar trend sequences, the horizontal dynamic merging is used to merge errors of adjacent time steps to optimize window division, and the asymmetric processing is used to process overestimated and underestimated errors to construct a precise confidence interval, which is stored in the recorder and dynamically updated.

[0180] In an embodiment, the system further includes a calibration unit configured to construct a precise confidence interval by vertically classifying to cluster similar trend sequences, horizontally dynamically merging errors of adjacent time steps to optimize window division, and asymmetrically processing overestimated and underestimated errors, store the confidence interval in the recorder, and dynamically update an error set.

[0181] In an embodiment, the calibration unit includes: The vertical segmentation subunit is configured to generate a plurality of clusters from historical prediction sequences by a clustering method, wherein each cluster contains sequences with similar trend and amplitude; the horizontal dynamic merging subunit is configured to define similarity of errors of adjacent time steps in each cluster, complete dynamic window division for each cluster, and make error distribution in each window consistent; the asymmetric error processing subunit is configured to define errors satisfying corresponding set conditions as overestimated and underestimated errors, calculate empirical quantiles of upper and lower limits of a confidence interval, and construct the confidence interval based on the empirical quantiles; the storage subunit is configured to store the clusters, corresponding dynamic windows, and confidence intervals in the recorder; and the dynamic updating subunit is configured to update error information in the recorder.

[0182] In an embodiment, the vertical segmentation subunit is configured to perform clustering analysis on historical prediction sequences by a k-means algorithm and a silhouette coefficient evaluation to determine an optimal number of clusters and assign labels; calculate similarity between a new prediction result and historical prediction sequences by soft dynamic time warping, select samples with required similarity, and assign a most frequent cluster to which the samples belong to the new prediction result; collect error items in the same cluster as the new prediction result, and construct an asymmetric prediction interval based on the error items.

[0183] In an embodiment, the vertical segmentation subunit is configured to collect error items of all samples in each time step in a cluster to which a new prediction result belongs to form an error subset; and construct an asymmetric prediction interval for the new prediction result based on error distribution in the error subset.

[0184] In an embodiment, the horizontal dynamic merging subunit is configured to compare error subsets of adjacent time steps in chronological order for each error term in a cluster, calculate statistical differences between the error subsets of adjacent time steps by a Kolmogorov-Smirnov test, and generate a p value; traverse from the first time step to the second last time step, and determine whether to merge adjacent time steps into a window according to the p value; if the p value exceeds a preset threshold, merge error terms of the current and next time steps; if the p value does not exceed the preset threshold, save error terms of an existing window, and start a new window from the current time step, and force to save the last window when reaching the end of the sequence; integrate error terms of all time steps of each window to form a merged error set, and construct a confidence interval of a time step corresponding to the window with the merged error set.

[0185] In an embodiment, the asymmetric error processing subunit is configured to separate error terms of each time step in a direction, and divide the error terms into overestimation errors and underestimation errors; calculate empirical quantiles of an overestimation error set and an underestimation error set to obtain empirical quantiles of upper and lower limits of a confidence interval; Construct an asymmetric confidence interval based on the empirical quantiles of the upper and lower limits of the confidence interval.

[0186] In an embodiment, the dynamic updating subunit is configured to calculate error terms between a new prediction result and a true value, and add the error terms to error subsets of corresponding clusters and windows according to a vertical classification rule; and automatically remove the oldest error term when a size of the error subset in a window exceeds a preset maximum capacity.

[0187] It should be noted that the specific implementation process of the above-mentioned multi-step time series prediction system 300 based on double-split conformal prediction and each unit can be clearly understood by those skilled in the art, which can be referred to the corresponding description in the foregoing method embodiments. For the convenience and brevity of description, it will not be repeated here.

[0188] The above-mentioned multi-step time series prediction system 300 based on double-split conformal prediction can be implemented in the form of a computer program, which can run on a computer device as shown in the computer device. Figure 6

[0189] Please refer to Figure 6 , Figure 6 is a schematic block diagram of a computer device provided by an embodiment of the present application. The computer device 500 can be a server, wherein the server can be a stand-alone server or a server cluster composed of multiple servers.

[0190] Please refer to Figure 6 ​The computer device 500 includes a processor 502, a memory, and a network interface 505 connected through a system bus 501, wherein the memory can include a non-volatile storage medium 503 and an internal memory 504.

[0191] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions which, when executed, can cause the processor 502 to perform a multi-step time series prediction method based on double-split conformal prediction.

[0192] The processor 502 is configured to provide computing and control capabilities to support the operation of the entire computer device 500.

[0193] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503, which, when executed by the processor 502, can cause the processor 502 to perform a multi-step time series prediction method based on double-split conformal prediction.

[0194] The network interface 505 is configured to perform network communication with other devices. Those skilled in the art can understand that the structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device 500 to which the scheme of the present application is applied. The specific computer device 500 can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components. Figure 6

[0195] The processor 502 is configured to run the computer program 5032 stored in the memory to implement all steps of the multi-step time series prediction method based on double-split conformal prediction.

[0196] It should be understood that, in the embodiments of the present application, the processor 502 can be a central processing unit (CPU), and the processor 502 can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), ready programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0197] ​Those skilled in the art can understand that all or part of the processes in the method of the above embodiments can be completed by instructing the relevant hardware by a computer program. The computer program includes program instructions, and the computer program can be stored in a storage medium, which is a computer readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the above method embodiments.

[0198] Therefore, the application further provides a storage medium. The storage medium can be a computer readable storage medium. The storage medium stores a computer program, wherein the computer program is executed by a processor to make the processor execute all steps of the multi-step time series prediction method based on double partition conformal prediction.

[0199] The storage medium can be a U disk, a mobile hard disk, a read-only memory (ROM), a magnetic disk or an optical disk, and various computer readable storage media that can store program codes.

[0200] Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been described in general terms in the above description. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the application.

[0201] In several embodiments provided by the application, it should be understood that the disclosed system and method can be implemented in other ways. For example, the system embodiments described above are only schematic. For example, the division of each unit is only a logical function division, and actual implementation can have another division manner. For example, a plurality of units or components can be combined or integrated into another system, or some features can be omitted or not executed.

[0202] The steps in the method of the embodiments of the application can be adjusted, combined and deleted in sequence according to actual needs. The units in the system of the embodiments of the application can be combined, divided and deleted according to actual needs. In addition, each functional unit in each embodiment of the application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit.

[0203] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a storage medium. Based on such an understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art, or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application.

[0204] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A multi-step time series forecasting method based on double-split conformal prediction, characterized in that, The method comprises: acquiring new input data; allocating the new input data to corresponding clustering clusters to obtain a clustering result; constructing a prediction interval from information extracted from a recorder according to the clustering result; adjusting corresponding content according to the prediction interval; wherein, by vertically classifying and clustering similar trend sequences, horizontally dynamically merging errors of adjacent time steps to optimize window division, and asymmetrically processing overestimated and underestimated errors, a precise confidence interval is constructed, stored in the recorder, and the error set is dynamically updated.

2. The multi-step time series forecasting method based on double-split conformal prediction according to claim 1, wherein, The method of constructing a precise confidence interval by vertically classifying and clustering similar trend sequences, horizontally dynamically merging errors of adjacent time steps to optimize window division, and asymmetrically processing overestimated and underestimated errors, storing in the recorder, and dynamically updating the error set comprises: generating a plurality of clusters from historical prediction sequences by a clustering method, wherein each cluster contains sequences with similar trends and amplitudes; defining the similarity of errors of adjacent time steps in the same cluster for each time step in each cluster, and completing dynamic window division for each cluster, wherein the error distribution in each window satisfies consistency; defining errors that meet corresponding set conditions as overestimated errors and underestimated errors, respectively calculating the empirical quantiles of the upper and lower limits of the confidence interval, and constructing the confidence interval based on the empirical quantiles of the upper and lower limits of the confidence interval; storing the clusters, corresponding dynamic windows, and confidence intervals in the recorder; updating error information in the recorder.

3. The multi-step time series forecasting method based on double-split conformal prediction of claim 2, wherein, The method of generating a plurality of clusters from historical prediction sequences by a clustering method comprises: performing clustering analysis on historical prediction sequences by k-means algorithm and silhouette coefficient evaluation to determine the optimal number of clusters and assign labels; calculating the similarity between new prediction results and historical prediction sequences using soft dynamic time warping, selecting samples with required similarity, and assigning the most frequent cluster to which the samples belong to the new prediction results; collecting error items in the same cluster as the new prediction results, and constructing an asymmetric prediction interval based on the error items.

4. The multi-step time series forecasting method based on double-split conformal prediction of claim 3, wherein, The method of collecting error items in the same cluster as the new prediction results, and constructing an asymmetric prediction interval based on the error items comprises: for the cluster to which the new prediction results belong, collecting error items of all samples in the cluster at each time step to form an error subset; constructing an asymmetric prediction interval for the new prediction results based on the error distribution in the error subset.

5. The multi-step time series forecasting method based on double-split conformal prediction of claim 4, wherein, The method of defining the similarity of errors of adjacent time steps in the same cluster for each time step in each cluster, and completing dynamic window division for each cluster comprises: for each error item in each cluster, compare the error subsets of adjacent time steps in chronological order, calculate the statistical difference between the error subsets of adjacent time steps by Kolmogorov-Smirnov test, and generate a p-value; starting from the first time step and traversing to the second last time step, determine whether to merge adjacent time steps into a window according to the p-value; if the p-value exceeds a preset threshold, merge the error items of the current and next time steps; if the p-value does not exceed the preset threshold, save the error items of the existing window, and start a new window from the current time as the starting point, and save the last window when reaching the end of the sequence; Integrate all error terms of each window to form a merged error set, and construct a confidence interval of the time step corresponding to the window with the merged error set.

6. The multi-step time series forecasting method based on double-split conformal prediction of claim 5, wherein, The definition meets the error terms corresponding to the set conditions, which are overestimated errors and underestimated errors, and the empirical quantiles of the upper and lower limits of the confidence interval are calculated, and the confidence interval is constructed based on the empirical quantiles of the upper and lower limits of the confidence interval, including: Directional separation of error terms for each time step, separating error terms into overestimated errors and underestimated errors; Calculate the empirical quantiles of the overestimated error set and the underestimated error set to obtain the empirical quantiles of the upper and lower limits of the confidence interval; Based on the empirical quantiles of the upper and lower limits of the confidence interval, construct an asymmetric confidence interval.

7. The multi-step time series forecasting method based on double-split conformal prediction of claim 6, wherein, The error information in the recorder is updated, including: Calculate the error term between the new prediction result and the true value, and add the error term to the error subset of the corresponding cluster and window according to the vertical classification rule; When the size of the error subset in the window exceeds the preset maximum capacity, the oldest error term is automatically removed.

8. The multi-step time series forecasting method based on double-split conformal prediction of claim 1, wherein, The new input data includes historical weather data and power generation, and the prediction interval includes the prediction interval of multi-step time series data.

9. The method of claim 1, wherein, The new input data includes server load, environmental temperature and cooling system data, and the prediction interval includes the prediction interval of multi-step load and temperature.

10. A multi-step time series prediction system based on bi-segment conformal prediction, characterized in that, Including: An acquisition unit is configured to acquire new input data; An allocation unit is configured to allocate the new input data to the corresponding clustering cluster to obtain a clustering result; An extraction unit is configured to extract information from the recorder to construct a prediction interval according to the clustering result; An adjustment unit is configured to adjust the corresponding content according to the prediction interval; By vertically classifying similar trend sequences and horizontally dynamically merging adjacent time step errors to optimize window division, asymmetrically processing overestimated and underestimated errors to construct accurate confidence intervals, and storing them in the recorder and dynamically updating error sets.

Images