Method and device for optimizing prediction performance of AI time series

Abstract

The invention relates to the technical field of data processing, and provides an AI time sequence prediction performance optimization method and device. The method comprises the following steps: horizontally splitting a to-be-predicted data set according to a time dimension to obtain a plurality of sub-data blocks; sending the plurality of sub-data blocks to a message bus; the message bus allocates the plurality of sub-data blocks to a plurality of prediction nodes; each prediction node obtains the distributed sub-data blocks so as to predict and obtain sub-results, and the sub-results are sent to the message bus; the message bus combines the sub-results to obtain a prediction result; according to the method, the problem that the direct unified processing effect is poor due to the fact that report results needing to be predicted are extremely complex and have many fields is solved, efficient time sequence prediction is achieved, and practicability is high.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for optimizing the performance of AI time series prediction. Background Technology

[0002] Large network operators bear the heavy responsibility of ensuring the stable and efficient operation of networks, which requires them to predict future data for network element equipment indicators. These indicators cover multiple aspects, including performance and usage.

[0003] However, in practice, the report results that need to be predicted are extremely complex, involving numerous fields. Directly using existing technologies such as large models for prediction results in poor performance and cannot meet the stringent real-time requirements of prediction work. This puts operators in a dilemma in planning network resources and ensuring network quality, and the existing technology for predicting future data of indicators has poor practicality.

[0004] Therefore, overcoming the shortcomings of the existing technology is an urgent problem to be solved in this technical field. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and apparatus for optimizing the performance of AI time series prediction. The purpose is to solve the problem in the prior art that the report results to be predicted are extremely complex and have many fields, resulting in poor performance when directly and uniformly processed.

[0006] The present invention adopts the following technical solution: In a first aspect, the present invention provides a method for optimizing the performance of AI time series prediction, comprising: The dataset to be predicted is horizontally split along the time dimension to obtain multiple sub-data blocks; The plurality of sub-data blocks are sent to the message bus; the message bus then distributes the plurality of sub-data blocks to a plurality of prediction nodes. Each prediction node acquires the allocated sub-data block to predict a sub-result, and sends the sub-result to the message bus; The message bus merges the various sub-results to obtain the prediction result.

[0007] Furthermore, before splitting the dataset to be predicted horizontally according to the time dimension to obtain multiple sub-data blocks, the process also includes: Triples are constructed using the original data to obtain a univariate time series; wherein, the dimensions of the triples include the indicator name, network element identifier, and sampling granularity; Calculate the data density of the univariate time series to obtain a segmentable identifier; determine a segmentation mode for the univariate time series from fixed window, sliding window, and dynamic adjustment according to the data density to obtain a segmentation strategy code; determine the start and end timestamps of the univariate time series according to the sampling granularity. The data block of the univariate time series is generated using the indicator name, the network element identifier, the divisible identifier, the divisibility strategy code, and the start and end timestamps. The dataset to be predicted is constructed using multiple of the data blocks.

[0008] Furthermore, the step of horizontally splitting the dataset to be predicted according to the time dimension to obtain multiple sub-data blocks includes: According to the start and end timestamps, the divisible data blocks are divided into multiple data blocks to be tested by using the segmentation strategy encoding the corresponding segmentation mode. Anomaly detection is performed on the plurality of data blocks to be tested, and abnormal data blocks and non-abnormal data blocks are marked. The plurality of abnormal data blocks and the plurality of non-abnormal data blocks are used as the plurality of sub-data blocks.

[0009] Furthermore, the prediction nodes include GPU nodes and CPU nodes; the message bus is implemented based on a unified message subscription model, which includes data sharding and topic allocation. The message bus distributes the plurality of sub-data blocks to the plurality of prediction nodes, including: The multiple sub-data blocks are obtained from the data sharding allocation topic, and the multiple sub-data blocks are allocated to each prediction node according to the resource status of each prediction node; When a sub-data block requires GPU acceleration, and / or the computational cost of the sub-data block exceeds a preset threshold, the sub-data block is allocated to a GPU node; otherwise, the sub-data block is allocated to a CPU node.

[0010] Furthermore, the sub-data block also includes indicator types; The message bus further includes distributing the plurality of sub-data blocks to the plurality of prediction nodes: When the indicator type is numerical, the sub-data block is allocated to the GPU node for processing; When the indicator type is text, the sub-data block is allocated to a CPU node for multi-core parallel processing.

[0011] Furthermore, the GPU node runs independently on the GPU; the CPU node runs independently on the CPU; and the unified message subscription model also includes a model weight update topic. The method further includes: The GPU node and the CPU node obtain weight parameters from the model weight update topic to work collaboratively with the model through a unified message subscription.

[0012] Furthermore, the unified message subscription model also includes a prediction result topic; the data block also includes a composite ID; The message bus merges the various sub-results to obtain the prediction result, including: Obtain sub-results for each prediction node from the prediction result topic; Based on the combined ID and start and end timestamps corresponding to the sub-results, the various sub-results are merged into a time series report as the prediction result.

[0013] Furthermore, the prediction node uses the Lag-LLaMA model to predict the assigned sub-data blocks and obtain sub-results.

[0014] In a second aspect, the present invention also provides an apparatus for optimizing the performance of AI time series prediction, used to implement the method for optimizing the performance of AI time series prediction described in the first aspect, the apparatus comprising: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor for performing the AI ​​time series prediction performance optimization method described in the first aspect.

[0015] Thirdly, the present invention also provides a non-volatile computer storage medium storing computer-executable instructions that are executed by one or more processors to perform the AI ​​time series prediction performance optimization method described in the first aspect.

[0016] This invention horizontally splits the dataset to be predicted according to the time dimension, solving the problem that the report results to be predicted are extremely complex and have many fields, resulting in poor performance when directly processed uniformly. Furthermore, it distributes the split sub-data blocks to multiple prediction nodes through a message bus, and each prediction node uses a large model for prediction. Finally, the results are merged through the message bus, thereby achieving real-time prediction. This makes it convenient for operators to plan network resources, ensure network quality, and achieve efficient time series prediction, making it highly practical. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments of the present invention will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0018] Figure 1 This is a flowchart illustrating a method for optimizing AI time series prediction performance according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the overall process of a method for optimizing the performance of AI time series prediction provided in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating a specific example of the first type of indicator data provided in this embodiment of the invention; Figure 4 This is a schematic diagram illustrating a specific example of the second type of indicator data provided in this embodiment of the invention; Figure 5 This is a flowchart illustrating step 10 provided in an embodiment of the present invention; Figure 6 This is a flowchart illustrating another step 10 provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating a specific example of a sharding strategy selection process provided in an embodiment of the present invention; Figure 8 This is a flowchart illustrating step 20 provided in an embodiment of the present invention; Figure 9 This is a flowchart illustrating another step 20 provided in an embodiment of the present invention; Figure 10 This is a flowchart illustrating step 40 provided in an embodiment of the present invention; Figure 11 This is a schematic diagram of the architecture of an AI time series prediction performance optimization device provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, the terms "inner", "outer", "longitudinal", "lateral", "upper", "lower", "top", "bottom", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and do not require that this invention must be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0021] In this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0022] In this application, unless otherwise expressly specified and limited, the term "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral part; it can be a direct connection or an indirect connection through an intermediate medium. Furthermore, the term "coupled" can refer to an electrical connection that enables signal transmission.

[0023] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0024] Example 1: To solve the above problems, such as Figure 1 As shown, this embodiment of the invention provides a method for optimizing the performance of AI time series prediction, including: Step 10: Horizontally split the dataset to be predicted according to the time dimension to obtain multiple sub-data blocks.

[0025] The dataset to be predicted is a segment of raw, real performance report data from the network element device to be predicted. In one embodiment, the dataset to be predicted can be data extracted from a data stream at time intervals.

[0026] like Figure 2 As shown, the dataset to be predicted is obtained and horizontally split. The process of splitting the dataset to be predicted will be explained below and will not be repeated here. Theoretically, data can be infinitely divided according to the time dimension, without a fixed upper limit. That is, the data can be divided into arbitrarily small time periods as needed; "infinite" can be understood as the flexibility of the splitting granularity. The splitting granularity can be dynamically adjusted according to the data density, frequency of change, etc., so that very fine-grained splitting can be performed when needed. In practice, due to limitations in computing resources, storage capacity, and data accuracy, the granularity of splitting will have a practical limit, but the upper limit can be very small, close to the concept of "infinite".

[0027] Step 20: Send the plurality of sub-data blocks to the message bus; the message bus distributes the plurality of sub-data blocks to the plurality of prediction nodes.

[0028] Since the metrics to be predicted all have significant time correlations, this embodiment of the invention uses a large model to achieve the prediction. A practical application scenario for one method of AI time series prediction performance optimization according to this embodiment of the invention can be: deploying a large model across multiple distributed prediction nodes, which collaborate through a message bus (e.g., Kafka). Figure 2 As shown, in one embodiment, the multiple sub-data blocks obtained in step 10 (i.e., Figure 2 Data sets 1, 2, and N are sent to the message bus, which then distributes these sub-data blocks to the various prediction nodes.

[0029] Step 30: Each prediction node obtains the allocated sub-data block to predict the sub-result, and sends the sub-result to the message bus.

[0030] In one embodiment, each prediction node can use the Lag-LLaMA large model to perform predictions: the Lag-LLaMA large model is deployed on each prediction node, and when the prediction node obtains the sub-data block allocated to it by the message bus, it predicts the sub-results through the Lag-LLaMA large model and returns the predicted sub-results to the message bus.

[0031] Step 40: The message bus merges the various sub-results to obtain the prediction result.

[0032] like Figure 2 As shown, in one embodiment, the sub-results obtained from each prediction node in step 30 (i.e., Figure 2 The prediction results 1, 2, and N are sent to the message bus, where they are merged by the data merging node to obtain the final prediction result.

[0033] This invention horizontally splits the dataset to be predicted according to the time dimension, solving the problem that the report results to be predicted are extremely complex and have many fields, resulting in poor performance when directly processed uniformly. Furthermore, it distributes the split sub-data blocks to multiple prediction nodes through a message bus, and each prediction node uses a large model for prediction. Finally, the results are merged through the message bus, thereby achieving real-time prediction. This makes it convenient for operators to plan network resources, ensure network quality, and achieve efficient time series prediction, making it highly practical.

[0034] The following is a further description of the AI ​​time series prediction performance optimization method according to an embodiment of the present invention: First, let's explain the dataset to be predicted.

[0035] Different networks involve different network elements and devices, and applications support various user service processes, such as authentication, addressing, calls, and internet access. For example, in a call scenario between a 4G mobile terminal and a 5G mobile terminal, data is transmitted between the 4G access network, the 4G core network, the IP Multimedia Subsystem (IMS) domain, the 5G core network, and the 5G access network. This process includes multiple sub-processes: 4G mobile terminal access, authentication, dialing, finding the network where the 5G mobile terminal is located, notifying the network where the 5G mobile terminal is located, establishing a call channel, starting the conversation, and ending the conversation. Each sub-process contains multiple metrics.

[0036] Because most network elements in technologies prior to 4G were physical devices, with the development of virtualization technology, 5G network elements began to be deployed in cloud pools. The collected data is the data output by the cloud pool; the network element devices deployed to the cloud carry the location information of the corresponding network element in the cloud pool. The data source of the dataset to be predicted in this embodiment of the invention includes 4G network element metrics collected from traditional Object Storage Service (OSS), as well as network element metrics collected from cloud resource pools after deployment.

[0037] For example, Deep Packet Inspection (DPI) call detail records (CDRs) are obtained by copying traffic between two network elements and then parsing the information in that traffic. When performing quality aggregation using DPI CDRs, source / destination information, error codes, etc., are parsed. Based on success / failure, time, and other information, metrics for a specified time period can be statistically analyzed. On the network management logic side, this corresponds to the network element, specifically which processes and interfaces the network element supports, and what types of metrics are available for each interface. The results are then reported. After obtaining this information, the upper-layer platform statistically analyzes these metrics according to the required dimensions, such as analyzing the S metrics of the K interfaces of all Authentication Management Function (AMF) devices at the same time point. Figure 3 This shows a specific example of data collected from an AMF (Anti-Functional Component) device. Each indicator name represents a network element indicator that was collected. For example... Figure 4 The image shows specific examples of various collected, processed, and calculated indicator data tables.

[0038] The dataset to be predicted is a set of original, real performance reports of the network element devices to be predicted. It is not a simple report, but a dataset that includes business and network element performance indicators related to process logic.

[0039] The purpose of AI time series prediction in this embodiment of the invention is to continuously predict the possible state of indicators of each network element based on the latest indicator data, and then predict whether certain network elements or services will generate alarms or even fail in the future based on the algorithm, thereby achieving performance optimization.

[0040] To achieve the above objectives, traditional prediction schemes require the use of various normalization models. After normalizing the indicator data, the obtained normalization parameters are used to build fitting models or prediction equations. Then, training data is divided using a sliding window and periodically fed into machine learning models. Machine learning models include algorithms such as Long Short-Term Memory (LSTM). The original indicators to be predicted can be as many as hundreds, with complex relationships between them, requiring manual combination and decomposition to discover these relationships. However, indicators such as user numbers are affected by holidays and daily routines, and traditional prediction schemes using function fitting cannot accurately reflect these patterns. Some models, such as the Autoregressive Integrated Moving Average (ARIMA) model or Prophet, rely on stationarity or manual rules and cannot generalize from zero samples.

[0041] To address this issue, in one embodiment, in step 30, the prediction node of this embodiment uses the Lag-LLaMA model to predict the allocated sub-data blocks, obtaining sub-results; the reason for selecting the Lag-LLaMA model for prediction is explained below: Basic time series models aim to take time series data as input and encode it accordingly to capture time dependencies. Examples of basic time series models include LSTM models and Recurrent Neural Networks (RNN) models. The Lag-LLaMA model, however, utilizes lagged features of past values ​​in a time series to capture these dependencies. Time series exhibit temporal patterns between current and lagged values ​​and contain calendar-related information, such as a day of the week or a week of the month. Therefore, Lag-LLaMA uses the input time series data as lagged covariates. The input to the Lag-LLaMA model is a label of a univariate time series at a given time step, consisting of lagged time steps and static covariates. The input sequence is mapped to the hidden dimensions of the attention module within the decoder through a linear projection layer. After passing through multiple masked decoder layers, the model outputs the parameters of the predicted distribution for the next time step through a distribution head.

[0042] Therefore, embodiments of the present invention utilize continuous data of a specific performance indicator of a system over a period of time, and process it to facilitate input into the Lag-LLaMA model for prediction. Any time-series data input into the Lag-LLaMA model should contain three basic elements: start time, target data, and data frequency (e.g., data collection granularity).

[0043] However, the existing Lag-LLaMA model uses a "single-card batch" inference approach, which has the following problems: a) When the length of the input sequence is greater than 50,000 or the batch size is greater than 5,000, the video memory will be filled up at once, directly causing the graphics processing unit (GPU) to overflow (OOM).

[0044] b) If the batch size is forcibly reduced, the preprocessing will become a new bottleneck on the central processing unit (CPU) side due to Python's Global Interpreter Lock (GIL) and memory bandwidth limitations.

[0045] c) Existing technologies only offer a choice between "model parallelism" or "data parallelism", which cannot achieve "zero-wait synchronization between chips" in a hybrid CPU and GPU cluster.

[0046] To address this problem, embodiments of the present invention employ data block and data partitioning strategies and a distributed data scheme to achieve usable and efficient time series forecasting. In one embodiment, such as... Figure 5 As shown, prior to step 10, the procedure further includes: Step 101a: Construct triples using the original data to obtain a univariate time series; wherein the dimensions of the triples include the indicator name, network element identifier, and sampling granularity.

[0047] To address the performance bottleneck of the Lag-LLaMA large model when processing network element device indicators, this invention provides a horizontally segmentable data block to facilitate subsequent integration with dynamic sharding strategies for efficient prediction.

[0048] A specific example of raw data is as follows: "2025-12-12 00:00:00 Network Element 1 Indicators ABCDE ..... 2025-12-12 00:00:01 Network Element 2 Indicators ABCDE ..... 2025-12-12 00:00:02 Network Element 1 Indicators ABCDE ..... 2025-12-12 00:00:03 Network Element 2 Indicators ABCDE....." In this embodiment of the invention, the original data is expanded into triplets of "indicator name - network element identifier - collection granularity" to obtain several univariate time series. Here, a univariate refers to a single indicator, represented by its name in the triplets. Each triplet represents a set of data for a specific indicator of a network element (represented by its network element identifier) ​​at the corresponding collection granularity; that is, the specific indicator data for a specific network element obtained after filtering. The collection granularity is, for example, 1 minute.

[0049] In one embodiment, before processing the original data, the original data can be standardized to a fixed time frequency, and the processed original data can be used to perform the operation in step 101a to solve the problems of uneven intervals and missing points that may exist in the original data.

[0050] Step 102a: Calculate the data density of the univariate time series to obtain a segmentable identifier; determine one of the segmentation modes for the univariate time series from fixed window, sliding window and dynamic adjustment according to the data density to obtain the segmentation strategy code; determine the start and end timestamps of the univariate time series according to the sampling granularity.

[0051] In one embodiment, the standard deviation of each univariate time series is calculated to measure its volatility. If the standard deviation is less than a preset threshold, the univariate time series is considered relatively stable and can be segmented; otherwise, the volatility is too high and it is not suitable for segmentation, thus determining the segmentation identifier of the univariate time series. A segmentation mode is determined from fixed window, sliding window, and dynamic adjustment according to the calculated standard deviation; the specific method of determining the segmentation mode is determined by those skilled in the art based on the specific application scenario. The start and end timestamps are the start time (e.g., 2025-12-12 00:00:00) and end time (e.g., 2025-12-12 00:00:03) of the univariate time series. In an optional embodiment, the inter-block overlap ratio between data blocks can also be calculated to ensure the continuity of out-of-distribution segments.

[0052] Step 103a: Use the indicator name, the network element identifier, the divisible identifier, the divisibility strategy code, and the start and end timestamps to generate the data block of the univariate time series.

[0053] In this embodiment of the invention, a data chunk (DataChunk) is created for each univariate time series, that is, a DataChunk is instantiated for each univariate series. The implementation method and format of the data chunk are determined by those skilled in the art based on the specific use case, and can be implemented using a custom data structure. A specific example of a data chunk is as follows: class DataChunk: chunk_id: str# Composite ID (including: dataset ID, timestamp start and end, and chunk number) splittable_flag: bool # Splittable flag: dynamically calculated based on data density (e.g., if the standard deviation is less than a threshold, it can be determined as splittable). time_range: tuple # Start and end timestamps: Start and end timestamps accurate to milliseconds (Note: ISO8601 format can be used) data_source: str#Network element device identifier (e.g., "BSC-01-192.168.1.1") sharding_strategy: int#Sharding strategy code (e.g., use "1" for a fixed window, "2" for a sliding window, and "3" for dynamic adjustment) overlap_rate: float # The percentage of overlap between blocks (e.g., 0.1 means 10% overlap). checksum: str# Hash checksum based on the data content data: array# Structured time series data (including: numerical indicators + category labels) The specific format and content of the indicator name, sampling granularity, network element identifier, divisible identifier, divisibility strategy code, and start and end timestamps shall be determined by those skilled in the art based on the specific application scenario, and are not limited here. The combined ID is a unique identifier for the data block. Both the combined ID and the hash checksum are used to uniquely identify each data block and verify data integrity.

[0054] The following is a specific example of converting univariate time series data into data blocks: “split_to_chunks(raw_df, grain, max_len, strategy, overlap): / / 1. Resampling → Unified granularity df = raw_df.set_index('timestamp').resample(grain).ffill() / / 2. Expand into single variables ts_list = [] for col in metric_columns: # Traverse M metrics for ne in df['ne_id'].unique(): # Traverse all network elements sub = df[df['ne_id']==ne][col].dropna().asfreq(grain, method='pad') ts_list.append( (ne, col, sub.values) ) / / 3. Generate DataChunk for each single variable sequence chunks = [] for (ne, metric, vals) in ts_list: std = np.std(vals) splittable = std < THRESHOLD # System configuration if splittable: step = int(max_len × (1 - overlap)) for start in range(0, len(vals)-max_len+1, step): end = start + max_len chunk_data = vals[start:end] chunk_id = f"{ne}{metric}{grain}{start}{end}" cksum = sha256(chunk_data.tobytes()).hexdigest() chunks.append( DataChunk(chunk_id, True, (time[start], time[end-1]), ne, strategy, overlap, cksum, chunk_data) ) else: # Not splittable, the whole segment is taken as one chunk chunks.append( DataChunk(f"{ne}_{metric}_full", False, (time[0], time[-1]), ne, 1, 0.0, sha256(vals.tobytes()).hexdigest(), vals) ) return chunks Wherein, raw_df represents the original data, a formatted multi-row, multi-column data table; grain represents the sampling granularity, in str format, such as "1min"; max_len represents the maximum processing length, in int format; overlap represents the inter-block overlap ratio, in float format, with a default of 0.1; strategy represents the segmentation strategy encoding, in int format, with three segmentation modes: fixed window, sliding window, and dynamic adjustment. In one embodiment, based on the single-floating-point characteristic of the σ / L ratio, the strategy encoding is automatically mapped to 1, 2, or 3 to achieve zero-parameter adaptiveness; in an optional embodiment, the segmentation strategy encoding calculation method can be: σ = np.std(ts) L = len(ts) R = σ / (L + 1e-6) return 1 if R<0.001 else (2 if R<0.01 else 3)” Step 104a: Construct the dataset to be predicted using multiple data blocks.

[0055] Collect all data blocks corresponding to the original data to be predicted, thus obtaining the dataset to be predicted used in step 10. After obtaining the dataset to be predicted, in one embodiment, such as... Figure 6 As shown, step 10 includes: Step 101b: According to the start and end timestamps, the divisible data blocks are divided into multiple data blocks to be tested by using the corresponding division pattern encoded by the division strategy.

[0056] Here, a divisible data block refers to a data block whose divisibility flag is true.

[0057] The data horizontal partitioning in this embodiment of the invention involves dividing the data into pieces according to the time series dimension. Each piece is then used to form an independent computational unit according to the data block format described in step 103a above, resulting in the data block to be tested. For example... Figure 7As shown, the dynamic segmentation strategy (i.e., a three-level segmentation mechanism) of this invention performs horizontal segmentation of the dataset to be predicted: the first level is a fixed window based on physical time, the second level is adaptive segmentation based on data features, and the third level is special handling of abnormal data. For example, the fixed time window can be 30 minutes, that is, according to the start and end timestamps in the preset format, the data of every 30 minutes is segmented into the same data block to be tested; the start and end timestamps are used to determine the start and end time windows of the segmentation. In one embodiment, the segmentation can be implemented using a segmentation function, in which the start and end timestamps are used in the segmentation function for sliding window or fixed window segmentation logic; then, a sliding window is used for dynamic adjustment; wherein, the segmentation strategy identifier is used to control whether the segment can continue to be split, ensuring that the granularity of the segmented data matches the model capacity. In an optional embodiment, the segmentation strategy encoding and the inter-block overlap ratio can be combined to specify how each segment is segmented and overlaps with adjacent segments, avoiding boundary effects from affecting prediction accuracy.

[0058] In one embodiment, the specific segmentation process of step 101b can be as follows: First, traverse each data block in the data stream and check whether the currently processed data block has a segmentable flag to determine whether the current data block needs to be further segmented. If the data block needs to be segmented, it is segmented according to the segmentation strategy encoding and the overlap ratio between blocks through the time dimension to generate the corresponding data block to be tested.

[0059] After horizontal partitioning, the data blocks to be tested are organized according to the time dimension: for example, the indicator values ​​of multiple indicators corresponding to multiple network element IDs at the times of "2024 / 8 / 28 4:15:00", "2024 / 8 / 28 2:30:00" and "2024 / 8 / 28 2:15:00".

[0060] In one embodiment, the dataset to be predicted is horizontally split based on the strength and type of field associations in the data block. A specific example of field association strength and type is shown in Table 1 below.

[0061] Table 1. Field Association Strength and Type

[0062] The composite ID includes a dataset ID, timestamp start and end dates, and fragment number. These fields have a direct logical relationship and collectively identify a data fragment. The segmentation strategy identifier describes the specific logic of the segmentation strategy and is directly related to the segmentation strategy code and the inter-block overlap ratio. The start and end timestamps define the time range of the data and are directly related to the timestamp start and end dates in the composite ID, representing the time interval directly. The network element device identifier identifies the device from which the data originates and is indirectly related to the data content, but not directly dependent on it. The segmentation strategy code specifies the type of segmentation strategy through encoding and is directly related to the segmentation strategy identifier and the inter-block overlap ratio. The inter-block overlap ratio describes the degree of overlap between blocks during segmentation and is directly related to the segmentation strategy identifier and the segmentation strategy code. The hash checksum is used to verify the integrity and accuracy of the data and is indirectly related to the structured time-series data, but not directly dependent on it. The structured time-series data is the core content of the data and is directly related to the hash checksum and timestamps, constituting the main part of the data.

[0063] This invention constructs data blocks, enabling the data to be predicted to be infinitely horizontally partitioned along the time series dimension. This allows large-scale datasets to be divided into multiple smaller batches, each of which can be processed independently on different CPU or GPU cores, thereby improving parallel processing efficiency. A pseudocode example of horizontal partitioning is provided below: for chunk in data_stream: if chunk.splittable_flag: sub_chunks = split_by_time(chunk, strategy=chunk.sharding_strategy,overlap=chunk.overlap_rate) dispatch_to_worker(sub_chunks) Each sub-data block can be independently allocated to a CPU or GPU node, enabling true parallel processing.

[0064] Step 102b: Perform anomaly detection on the plurality of data blocks to be tested, mark abnormal data blocks and non-abnormal data blocks, and use the plurality of abnormal data blocks and the plurality of non-abnormal data blocks as the plurality of sub-data blocks.

[0065] The specific methods for anomaly detection and marking are selected by those skilled in the art based on the specific application scenario and are not limited herein. This invention provides an anomaly fragment hot-processing mechanism: data blocks with data fluctuations exceeding thresholds, missing fields, and / or computational delays are identified as anomaly data fragments, which are then marked and dynamically rescheduled. In one embodiment, during subsequent processing, the anomaly data fragments can trigger GPU-priority computation or cross-node migration processing; by automatically identifying and hot-processing anomaly fragments, the system robustness and prediction real-time performance are improved.

[0066] In one embodiment, a weighted fusion or sliding window smoothing strategy is employed for data in overlapping regions to ensure prediction accuracy. Weighted fusion involves assigning different weights to the data in overlapping regions and then performing weighted averaging or other weighting operations to fuse information from different data blocks. The sliding window smoothing strategy involves defining a fixed-size sliding window containing a certain number of data points, calculating the average value of the data within each sliding window position, and then using these average values ​​as the smoothed data points. This embodiment of the invention combines a segmentation strategy with an incremental merging mechanism to achieve online reconstruction of high-precision prediction results.

[0067] In one embodiment, a specific example of a dynamic block partitioning strategy is as follows: “def dynamic_split(data_stream, model_capacity): # Level 1: Fixed window based on physical time base_chunks = time_window_split(data_stream, delta='30min') # Level 2: Adaptive Segmentation Based on Data Features for chunk in base_chunks: if should_split(chunk, model_capacity): # Use sliding window for dynamic adjustment sub_chunks = sliding_window_split( chunk, window_size=model_capacity×0.8, overlap = 0.15 × model_capacity ) update_chunk_list(sub_chunks) # Level 3: Special Handling of Abnormal Data mark_hotspot_chunks(chunk_list)” like Figure 7 As shown in the illustration, this embodiment of the invention also provides a specific schematic diagram of a data segmentation strategy selection process. First, the key features of the currently processed data are identified; if the data has strong temporal correlation, it enters the horizontal segmentation process; if the data has multi-indicator correlation, it enters the vertical segmentation process. Specifically, for data with strong temporal correlation, horizontal segmentation is performed, that is, segmenting according to the time dimension following steps 101b to 102b.

[0068] In one embodiment, the present invention also provides a global load feedback and adaptive optimization scheme: the coordination controller dynamically adjusts the sharding strategy based on the completion status, throughput and latency indicators of each prediction node: when the prediction node is under high load, the sharding granularity automatically increases to reduce communication overhead; when the prediction node is idle, the sharding granularity decreases to increase parallelism; by forming a closed-loop optimization of sharding strategy and distributed scheduling, the overall system performance is adaptively improved.

[0069] In one embodiment, the prediction nodes include GPU nodes and CPU nodes; the message bus is implemented based on a unified message subscription model, which includes data sharding and topic allocation. The unified message subscription model also includes a model weight update topic. This embodiment of the invention distributes the prediction nodes loading the Lag-LLaMA model, allowing each prediction node to run independently on different CPU cores or GPU cores, and to collaborate through the unified message subscription model, thereby fully utilizing resources. Here, a message bus is a software architecture pattern used to transmit messages or data between different applications or components; the message bus allows each component to send and receive messages independently. For example, a unified message queue is an implementation of a message bus, a message processing system that allows multiple producers and consumers to communicate through a shared queue.

[0070] In one embodiment, the present invention is implemented based on a unified message subscription model. A unified message subscription model is a messaging pattern in which a publisher sends messages to one or more topics, and subscribers can subscribe to these topics to receive messages of interest. A single message can be received by multiple subscribers. A topic is a logical channel used to categorize messages. A publisher sends messages to specific topics, and subscribers can subscribe to these topics to receive messages. Subscription refers to the process by which a subscriber registers with one or more topics to receive messages on those topics.

[0071] The unified message subscription model in this embodiment of the invention includes three core topics: `data_shard_assign`: shard allocation, recording the `chunk_id` of each shard and the allocation node; `model_weights_update`: model parameter synchronization, ensuring that distributed nodes use a consistent model; and `prediction_result`: prediction result collection, used for incremental stream merging. Figure 8 As shown, in step 20, the message bus allocates the plurality of sub-data blocks to the plurality of prediction nodes, including: Step 201a: Obtain the multiple sub-data blocks from the data sharding allocation topic, and allocate the multiple sub-data blocks to the prediction nodes according to the resource status of each prediction node.

[0072] The data sharding allocation topic records the prediction nodes and their status allocated to each shard (i.e., sub-data block). In one embodiment, the coordination controller continuously monitors the resource status of each prediction node (e.g., number of CPU cores, GPU availability, and current load); each horizontally partitioned sub-data block carries complete field information (i.e., combined ID, start and end timestamps, partitioning strategy code, and inter-block overlap ratio), assisting the controller in intelligently matching suitable nodes to achieve dynamic sharding allocation and resource-aware scheduling. In one embodiment, the coordination controller reads the data sharding allocation topic and allocates shards according to the resource status of each node (e.g., CPU or GPU availability, load, etc.).

[0073] Step 202a: When a sub-data block requires GPU acceleration, and / or the computational load of the sub-data block exceeds a preset threshold, the sub-data block is allocated to a GPU node; otherwise, the sub-data block is allocated to a CPU node.

[0074] The preset threshold is selected by those skilled in the art based on the specific application scenario and is not limited here. Through dynamic allocation, computationally intensive shards that rely on GPU acceleration are prioritized for allocation to GPU nodes, while lightweight shards are processed in parallel on CPU multi-cores. In this way, large-scale test datasets can be split into multiple independently processable small batches; each batch can be allocated in parallel to different CPU or GPU nodes for computation; the shard granularity and overlap rate can be dynamically adjusted to balance parallel efficiency and prediction accuracy. This embodiment of the invention constructs data blocks to achieve horizontal partitioning of the dataset to be predicted in the time dimension, so as to facilitate the allocation of sub-data blocks to various prediction nodes for prediction, thereby achieving adaptive matching between shard granularity and node resources, solving the performance bottleneck under large-scale complex reports when using the Lag-LLaMA large model for prediction.

[0075] This invention also provides a specific example of an allocation process: "def infinite_horizontal_split(data_flow, max_workers): # Initialize the shard queue chunk_queue = CircularBuffer(size=max_workers×3) # Dynamic Fragmentation Executor while data_flow.has_more(): current_window = data_flow.next_window( size = adaptive_window_size(), overlap = current_overlap() ) # GPU resource awareness allocation if requires_gpu_acceleration(current_window): assign_to_gpu_node(current_window) else: # CPU multi-core parallel processing parallel_split(current_window, chunk_queue) # Real-time status synchronization (sent to a distributed message queue) publish_status_update({ 'processed': chunk_queue.completed_count, 'throughput': calculate_throughput() })” In one embodiment, the GPU node runs independently on the GPU; the CPU node runs independently on the CPU; in one embodiment, the sub-data block further includes an indicator type; In step 20, as Figure 9 As shown, the message bus further includes the following to allocate the plurality of sub-data blocks to the plurality of prediction nodes: Step 201b: When the indicator type is numerical, the sub-data block is allocated to the GPU node for processing.

[0076] like Figure 7 As shown, for data with multiple correlation indicators, vertical partitioning is performed, that is, grouping according to different indicators. If the indicator is numerical, GPU sharding is used for processing.

[0077] Step 202b: When the indicator type is text, the sub-data block is allocated to the CPU node for multi-core parallel processing.

[0078] If the metric is text-based, CPU sharding is used for processing.

[0079] In the prior art, when splitting data, it is necessary to combine and split various types of data to input them into the model used for prediction (e.g., the Lag-LLaMA model in the embodiments of the present invention). Each prediction must be performed by dividing the data to be predicted into multiple groups, which results in low running efficiency, poor prediction performance, and is completely unsatisfactory. Furthermore, errors may occur due to the incompatibility between the allocation strategy of memory usage and the batch processing volume.

[0080] This invention addresses the shortcomings of existing technologies using the Lag-LLaMA model for prediction. It constructs data blocks and segments and distributes the data to be predicted according to these blocks, assigning each block to a different prediction node for prediction and then merging the prediction results. Because the original data is pre-organized by constructing data blocks, the data that would otherwise require grouping hundreds of fields in existing technologies can be horizontally segmented using a single time dimension. Since the purpose of this invention is to predict the state of the next time period using real-time collected data, horizontally segmenting the data to be predicted along the time dimension allows for simple and rapid allocation of collected data from the same time period to different prediction nodes. Furthermore, strategy allocation is implemented during variable segmentation, using these as input samples to optimize the prediction data in real time.

[0081] Specifically, because the embodiments of the present invention can allocate the collected data in the same time period to different prediction nodes for prediction, and the sub-data blocks obtained by segmenting according to the data block format also carry a segmentable identifier and a segmentation strategy encoding, the segmentable sub-data blocks can be further segmented according to the segmentation strategy encoding, and a series of partitions between GPUs or CPUs can be performed during allocation; therefore, it is not affected by the length of the input sequence of the dataset to be predicted, and there will be no situation where the video memory is filled up at once, directly causing an OOM, and thus preprocessing will not become a new bottleneck. By using the Lag-LLaMA model to predict at each prediction node, and then uniformly collecting the prediction results of each prediction node to form a prediction report, the data can be predicted in parallel on multiple prediction nodes. "Model parallelism" and "data parallelism" can be achieved simultaneously, so "zero-wait synchronization between chips" can be achieved in a CPU and GPU hybrid cluster.

[0082] This invention combines a design strategy for index segmentation encoding, an overlap ratio, and a data type for structured time series data. This enables the Lag-LLaMA model to achieve linear speedup on a hybrid cluster of ordinary CPUs and GPUs when dealing with terabyte (TB) level time series, without requiring modification of the original weights of the Lag-LLaMA model.

[0083] Following step 30, after each prediction node subscribes to the topic, it obtains the sub-data block it is responsible for, runs the Lag-LLaMA large model independently, inputs the data of the sub-data block into the Lag-LLaMA large model for prediction, and obtains the sub-result.

[0084] The method further includes: the GPU nodes and the CPU nodes obtaining weight parameters from the model weight update topic to collaborate with the model through unified message subscription. The model weight update topic distributes data through inter-node message exchange to ensure consistent model weights across distributed nodes and supports dynamic online fine-tuning. Asynchronous decoupling and high-availability collaboration of distributed prediction nodes are achieved through message queues, avoiding centralized scheduling bottlenecks.

[0085] In one embodiment, asynchronous collaboration driven by a unified message queue is achieved based on a custom topic. All prediction nodes complete tasks such as sharding task acquisition, model parameter synchronization, and prediction result reporting by subscribing to a Kafka topic. The unified message subscription model also includes a prediction result topic, which is used to realize the return of sharded prediction results and achieve incremental merging.

[0086] like Figure 10 As shown, step 40 includes: Step 401: Obtain the sub-results of each prediction node from the prediction result topic.

[0087] In one embodiment, the sub-results of the completed prediction carry a group ID and are sent to the prediction result topic; each prediction result shard carries a group ID and a start and end timestamp, and the merging engine listens to the prediction result topic in real time.

[0088] Step 402: According to the combined ID and start and end timestamps corresponding to the sub-results, merge the sub-results into a time series report as the prediction result.

[0089] In one embodiment, the message bus merging engine reassembles the sub-results from each prediction node into a complete prediction result according to the order of the group IDs corresponding to the sub-results.

[0090] like Figure 11 The diagram shown is an architectural schematic of an AI time series prediction performance optimization device according to an embodiment of the present invention. The device for AI time series prediction performance optimization in this embodiment includes one or more processors 21 and a memory 22. Figure 11 Take a processor 21 as an example.

[0091] Processor 21 and memory 22 can be connected via a bus or other means. Figure 11 Taking the example of a connection between China and Israel via a bus.

[0092] The memory 22, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs and non-volatile computer-executable programs, such as the AI ​​time series prediction performance optimization method in Embodiment 1. The processor 21 executes the AI ​​time series prediction performance optimization method by running the non-volatile software programs and instructions stored in the memory 22.

[0093] Memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, memory 22 may optionally include memory remotely located relative to processor 21, which can be connected to processor 21 via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0094] The program instructions / modules are stored in the memory 22. When executed by one or more processors 21, they perform the AI ​​time series prediction performance optimization method described in Embodiment 1 above, for example, performing each step of the AI ​​time series prediction performance optimization method described above.

[0095] It is worth noting that the information interaction and execution process between the modules and units in the above-mentioned device and system are based on the same concept as the processing method embodiment of the present invention. For details, please refer to the description in the method embodiment of the present invention, and will not be repeated here.

[0096] Those skilled in the art will understand that all or part of the steps in the various methods of the embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc.

[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing the performance of AI time series prediction, characterized in that, include: The dataset to be predicted is horizontally split along the time dimension to obtain multiple sub-data blocks; Send the plurality of sub-data blocks to the message bus; The message bus distributes the plurality of sub-data blocks to the plurality of prediction nodes; Each prediction node acquires the allocated sub-data block to predict a sub-result, and sends the sub-result to the message bus; The message bus merges the various sub-results to obtain the prediction result.

2. The method for optimizing AI time series prediction performance according to claim 1, characterized in that, Before horizontally splitting the dataset to be predicted according to the time dimension to obtain multiple sub-data blocks, the following is also included: Triples are constructed using the original data to obtain a univariate time series; wherein, the dimensions of the triples include the indicator name, network element identifier, and sampling granularity; Calculate the data density of the univariate time series to obtain a segmentable identifier; determine a segmentation mode for the univariate time series from fixed window, sliding window, and dynamic adjustment according to the data density to obtain a segmentation strategy code; determine the start and end timestamps of the univariate time series according to the sampling granularity. The data block of the univariate time series is generated using the indicator name, the network element identifier, the divisible identifier, the divisibility strategy code, and the start and end timestamps. The dataset to be predicted is constructed using multiple of the data blocks.

3. The method for optimizing AI time series prediction performance according to claim 2, characterized in that, The step of horizontally splitting the dataset to be predicted according to the time dimension to obtain multiple sub-data blocks includes: According to the start and end timestamps, the divisible data blocks are divided into multiple data blocks to be tested by using the segmentation strategy encoding the corresponding segmentation mode. Anomaly detection is performed on the plurality of data blocks to be tested, and abnormal data blocks and non-abnormal data blocks are marked. The plurality of abnormal data blocks and the plurality of non-abnormal data blocks are used as the plurality of sub-data blocks.

4. The method for optimizing AI time series prediction performance according to claim 1, characterized in that, The prediction nodes include GPU nodes and CPU nodes; the message bus is implemented based on a unified message subscription model, which includes data sharding and topic allocation. The message bus distributes the plurality of sub-data blocks to the plurality of prediction nodes, including: The multiple sub-data blocks are obtained from the data sharding allocation topic, and the multiple sub-data blocks are allocated to each prediction node according to the resource status of each prediction node; When a sub-data block requires GPU acceleration, and / or the computational cost of the sub-data block exceeds a preset threshold, the sub-data block is allocated to a GPU node; otherwise, the sub-data block is allocated to a CPU node.

5. The method for optimizing AI time series prediction performance according to claim 4, characterized in that, The sub-data block also includes indicator types; The message bus further includes distributing the plurality of sub-data blocks to the plurality of prediction nodes: When the indicator type is numerical, the sub-data block is allocated to the GPU node for processing; When the indicator type is text, the sub-data block is allocated to a CPU node for multi-core parallel processing.

6. The method for optimizing AI time series prediction performance according to claim 4, characterized in that, The GPU node runs independently on the GPU; the CPU node runs independently on the CPU; the unified message subscription model also includes a model weight update topic; The method further includes: The GPU node and the CPU node obtain weight parameters from the model weight update topic to work collaboratively with the model through a unified message subscription.

7. The method for optimizing AI time series prediction performance according to claim 4, characterized in that, The unified message subscription model also includes prediction result topics; the data block also includes a composite ID; The message bus merges the various sub-results to obtain the prediction result, including: Obtain sub-results for each prediction node from the prediction result topic; Based on the combined ID and start and end timestamps corresponding to the sub-results, the various sub-results are merged into a time series report as the prediction result.

8. The method for optimizing AI time series prediction performance according to any one of claims 1-7, characterized in that, The prediction node uses the Lag-LLaMA model to predict the assigned sub-data blocks and obtain sub-results.

9. A non-volatile computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, which are executed by one or more processors to perform the AI ​​time series prediction performance optimization method according to any one of claims 1-8.

10. An apparatus for optimizing the performance of AI time series prediction, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the processor for performing the AI ​​time series prediction performance optimization method according to any one of claims 1-8.

Images