CN-121980171-A - AI time sequence prediction performance optimization method and device

CN121980171ACN 121980171 ACN121980171 ACN 121980171ACN-121980171-A

Abstract

The invention relates to the technical field of data processing, and provides a method and a device for optimizing AI time sequence prediction performance. The method comprises the steps of dividing a data set to be predicted according to a time dimension level to obtain a plurality of sub data blocks, sending the plurality of sub data blocks to a message bus, distributing the plurality of sub data blocks to a plurality of prediction nodes by the message bus, obtaining the distributed sub data blocks by each prediction node to predict to obtain a sub result, sending the sub result to the message bus, and merging the sub results by the message bus to obtain a prediction result.

Inventors

ZHAO HAIXIN
YE ZHIGANG

Assignees

武汉绿色网络股份有限公司

Dates

Publication Date: 20260505
Application Date: 20251231

Claims (10)

1. A method for AI time sequence predictive performance optimization, comprising: splitting a data set to be predicted according to a time dimension level to obtain a plurality of sub data blocks; the message bus distributes the plurality of sub data blocks to a plurality of prediction nodes; each prediction node acquires the allocated sub-data blocks to predict and obtain sub-results, and sends the sub-results to the message bus; And combining all the sub-results by the message bus to obtain a predicted result.
2. The method of AI time sequence prediction performance optimization of claim 1, further comprising, prior to said splitting of the data set to be predicted at the time dimension level to obtain a plurality of sub-data blocks: Constructing a triplet by using the original data to obtain a univariate time sequence, wherein the dimension of the triplet comprises an index name, a network element identifier and a sampling granularity; determining one of the segmentation modes for the univariate time sequence from a fixed window, a sliding window and dynamic adjustment according to the data density to obtain a segmentation strategy code; generating a data block of the univariate time sequence by using the index name, the network element identifier, the segmentable identifier, the segmentation strategy code and the start-stop time stamp; The data set to be predicted is constructed using a plurality of the data blocks.
3. The method of AI time sequence prediction performance optimization of claim 2, wherein splitting the data set to be predicted at the time dimension level to obtain a plurality of sub-data blocks comprises: according to the start-stop time stamp, the data blocks which can be segmented are segmented by using the segmentation mode corresponding to the segmentation strategy codes, and a plurality of data blocks to be detected are obtained; and performing anomaly detection on the data blocks to be detected, marking the anomalous data blocks and the non-anomalous data blocks, and taking the anomalous data blocks and the non-anomalous data blocks as the sub-data blocks.
4. The method of claim 1, wherein the prediction nodes comprise GPU nodes and CPU nodes, wherein the message bus is implemented based on a unified message subscription model, and wherein the unified message subscription model comprises a data slicing distribution theme; the message bus allocating the plurality of sub-data blocks to a plurality of prediction nodes comprises: acquiring the plurality of sub-data blocks from the data fragment distribution subject, and distributing the plurality of sub-data blocks to each prediction node according to the resource condition of each prediction node; And when the sub data block needs GPU acceleration and/or the calculated amount of the sub data block is larger than a preset threshold value, distributing the sub data block to a GPU node, and otherwise distributing the sub data block to a CPU node.
5. The method of AI time sequence prediction performance optimization of claim 4 wherein the sub-data block further includes an index type; the message bus allocating the plurality of sub-data blocks to a plurality of prediction nodes further comprises: when the index type is numerical, distributing the sub data blocks to GPU nodes for processing; and when the index type is text type, distributing the sub data blocks to CPU nodes for multi-core parallel processing.
6. The method of AI time sequence prediction performance optimization as recited in claim 4 wherein the GPU node is independently running on a GPU, the CPU node is independently running on a CPU, the unified message subscription model further comprises a model weight update theme; The method further comprises the steps of: the GPU node and the CPU node acquire weight parameters from the model weight updating theme so as to cooperatively work through a unified message subscription model.
7. The method of AI time sequence prediction performance optimization as recited in claim 4 wherein the unified message subscription model further comprises a prediction result topic, wherein the data block further comprises a combined ID; the message bus merges the sub-results, and the obtaining of the predicted result includes: Obtaining sub-results of each prediction node from the prediction result theme; And merging each sub-result into a time sequence report as a prediction result according to the combined ID and the start-stop time stamp corresponding to the sub-result.
8. The method of any one of claims 1-7, wherein the prediction node predicts the assigned sub-data blocks using a Lag-LLaMA model to obtain sub-results.
9. A non-transitory computer storage medium storing computer-executable instructions for execution by one or more processors for performing the AI time sequence prediction performance optimization method of any of claims 1-8.
10. An apparatus for AI time sequence prediction performance optimization, comprising: And a memory communicatively coupled to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executable by the processor for performing the AI time sequence prediction performance optimization method of any of claims 1-8.

Description

AI time sequence prediction performance optimization method and device Technical Field The invention relates to the technical field of data processing, in particular to a method and a device for optimizing AI time sequence prediction performance. Background The large network operators bear the heavy duty of guaranteeing the stable and efficient operation of the network, and for this purpose, future data prediction needs to be carried out on indexes of network element equipment. These indicators that need to be predicted cover a number of aspects, including performance and usage. However, in actual operation, report results to be predicted are extremely complex, the related fields are numerous, the performance of directly using large models and the like in the prior art for prediction is poor, the severe requirements of prediction work on real-time performance cannot be met, and therefore operators are dilemma in the aspects of planning network resources, guaranteeing network quality and the like, and the scheme of the prior art for performing index future data prediction is poor in practicality. In view of this, overcoming the drawbacks of the prior art is a problem to be solved in the art. Disclosure of Invention The invention aims to solve the problems that report results to be predicted are extremely complex and fields are numerous, so that direct unified processing effect is poor in the prior art. The invention adopts the following technical scheme: In a first aspect, the present invention provides a method for AI time sequence prediction performance optimization, comprising: splitting a data set to be predicted according to a time dimension level to obtain a plurality of sub data blocks; the message bus distributes the plurality of sub data blocks to a plurality of prediction nodes; each prediction node acquires the allocated sub-data blocks to predict and obtain sub-results, and sends the sub-results to the message bus; And combining all the sub-results by the message bus to obtain a predicted result. Further, before splitting the data set to be predicted according to the time dimension level to obtain a plurality of sub data blocks, the method further includes: Constructing a triplet by using the original data to obtain a univariate time sequence, wherein the dimension of the triplet comprises an index name, a network element identifier and a sampling granularity; determining one of the segmentation modes for the univariate time sequence from a fixed window, a sliding window and dynamic adjustment according to the data density to obtain a segmentation strategy code; generating a data block of the univariate time sequence by using the index name, the network element identifier, the segmentable identifier, the segmentation strategy code and the start-stop time stamp; The data set to be predicted is constructed using a plurality of the data blocks. Further, splitting the data set to be predicted according to the time dimension level to obtain a plurality of sub data blocks includes: according to the start-stop time stamp, the data blocks which can be segmented are segmented by using the segmentation mode corresponding to the segmentation strategy codes, and a plurality of data blocks to be detected are obtained; and performing anomaly detection on the data blocks to be detected, marking the anomalous data blocks and the non-anomalous data blocks, and taking the anomalous data blocks and the non-anomalous data blocks as the sub-data blocks. The message bus is realized based on a unified message subscription model, and the unified message subscription model comprises a data fragment distribution theme; the message bus allocating the plurality of sub-data blocks to a plurality of prediction nodes comprises: acquiring the plurality of sub-data blocks from the data fragment distribution subject, and distributing the plurality of sub-data blocks to each prediction node according to the resource condition of each prediction node; And when the sub data block needs GPU acceleration and/or the calculated amount of the sub data block is larger than a preset threshold value, distributing the sub data block to a GPU node, and otherwise distributing the sub data block to a CPU node. Further, the sub data block further includes an index type; the message bus allocating the plurality of sub-data blocks to a plurality of prediction nodes further comprises: when the index type is numerical, distributing the sub data blocks to GPU nodes for processing; and when the index type is text type, distributing the sub data blocks to CPU nodes for multi-core parallel processing. Further, the GPU node independently operates on the GPU, the CPU node independently operates on the CPU, and the unified message subscription model also comprises a model weight updating theme; The method further comprises the steps of: the GPU node and the CPU node acquire weight parameters from the model weight updating theme so as to cooperatively work