Multi-model parallel integrated adaptive time series prediction method and system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING LANGE CLOUD BUSINESS TECH CO LTD
- Filing Date
- 2026-02-06
- Publication Date
- 2026-06-05
Smart Images
Figure CN122153256A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of time series forecasting technology, and in particular to a multi-model parallel integrated adaptive time series forecasting method and system. Background Technology
[0002] In applications such as financial risk control, macroeconomic analysis, supply chain demand management, and energy load scheduling, time-series forecasting is a crucial foundational capability supporting business decisions and resource allocation. As business environments and data sources become increasingly complex, forecasting targets are often influenced by multiple factors simultaneously, such as industry indicators, macroeconomic variables, market prices, corporate operating data, and external events. The introduction of multi-source heterogeneous data necessitates that forecasting systems not only possess the ability to uniformly process data of different frequencies and calibers, but also continuously output stable and interpretable forecasting results within a short modeling cycle to meet the engineering requirements of rolling forecasting and rapid iteration.
[0003] In existing technologies, common approaches to time series forecasting include statistical model-based forecasting methods (such as exponential smoothing and ARIMA-like methods), machine learning-based regression forecasting methods (such as tree models and linear regression), and deep learning-based sequence modeling methods (such as RNN / LSTM / Transformer). In engineering practice, it is often necessary to perform frequency alignment, missing data handling, and indicator transformation on multi-source data, and construct time lag features, year-on-year / month-on-month features, etc., to improve modeling performance. Subsequently, different models are trained, parameters are tuned, and evaluated, and finally, the output results are combined or optimized. However, the above process usually has the following shortcomings: First, the data from multiple sources vary greatly in scope, frequency, and quality, making preprocessing and feature construction reliant on human experience and making it difficult to form a reusable standardized process. Noise and bias before the data enters the model are easily amplified. Second, multi-model training and rolling window prediction are often performed sequentially. Parameter tuning, retraining, and multi-task prediction bring significant computational overhead and cycle delays, making it difficult to balance prediction accuracy and update frequency with limited resources. Third, the output scale, stability, and error characteristics of different models vary significantly. Existing integration methods often rely on simple averaging or single-model selection, lacking a systematic measurement of model consistency and robustness, which leads to problems such as fluctuations in prediction results, insufficient controllability, and inconsistent output formats.
[0004] Therefore, in time series prediction driven by multi-source heterogeneous data, the insufficient standardization of data frequency alignment and feature construction, the low computational efficiency of multi-model training and rolling prediction, and the poor stability and consistency of integrated output have become problems that urgently need to be solved. Summary of the Invention
[0005] This application provides a multi-model parallel integrated adaptive time series prediction method and system, which aims to solve the problems of insufficient standardization of data frequency alignment and feature construction, low computational efficiency of multi-model training and rolling prediction, and poor stability and consistency of integrated output in the existing technology for time series prediction driven by multi-source heterogeneous data.
[0006] A first aspect is a multi-model parallel ensemble adaptive time series prediction method, the method comprising:
[0007] Obtain historical time-series data of the indicator to be predicted, and obtain multi-source heterogeneous data related to the indicator to be predicted;
[0008] Data processing is performed on the historical time-series data and multi-source heterogeneous data to obtain a unified time-series dataset for modeling;
[0009] Construct an input feature set based on the unified time-series dataset;
[0010] Configure a set of parallel computing tasks for the metric to be predicted. The set of parallel computing tasks includes parameter optimization tasks and multiple rolling window training and prediction tasks. Submit the set of parallel computing tasks to the job scheduling system for parallel execution.
[0011] Train and / or invoke multiple prediction models in various parallel computing tasks to obtain prediction sequences corresponding to each prediction model;
[0012] Parallel integration is performed on the prediction sequences corresponding to each prediction model to obtain an integrated prediction sequence;
[0013] An adaptive trend correction is performed on the integrated prediction sequence to obtain the final prediction sequence;
[0014] The final predicted sequence is output and stored in a structured manner.
[0015] Optionally, in the above scheme, the data processing includes: processing missing values, outliers, and frequency alignment on the historical time-series data and multi-source heterogeneous data to form the unified time-series dataset with unified time granularity and unified time index.
[0016] Optionally, the multi-source heterogeneous data mentioned above may include at least one or more of the following: macroeconomic indicator data, industry indicator data, financial market indicator data, commodity market indicator data, and enterprise operating indicator data.
[0017] Optionally, in the above scheme, the construction of the input feature set includes time delay features, which are composed of lag terms generated from the unified time series dataset according to a preset lag order.
[0018] Optionally, in the above scheme, the method further includes performing data transformation on the unified time series dataset, the data transformation including normalization transformation and / or standardization transformation.
[0019] Optionally, the above scheme further includes year-on-year conversion processing: converting the original numerical sequence into a year-on-year growth rate sequence; performing predictive modeling based on the year-on-year growth rate sequence to obtain a predicted year-on-year growth rate value; and performing an inverse conversion on the predicted year-on-year growth rate value to obtain a numerical prediction value.
[0020] In the above scheme, optionally, the rolling window training prediction task corresponds to different time series parameter configurations, and the time series parameters include at least the rolling window size parameter and the test dataset size parameter.
[0021] Optionally, the various prediction models mentioned above include: temporal convolutional network models, Transformer models, graph attention network models, gradient boosting tree models, and neural network regression models.
[0022] Optionally, in the above scheme, the parallel integration includes: performing consistency calculation on the prediction sequences output by different prediction models based on a consistency metric, and determining integration rules based on the consistency calculation results to generate the integrated prediction sequence;
[0023] The adaptive trend correction includes: determining a reference trend sequence based on historical time series data, calculating the trend deviation between the integrated prediction sequence and the reference trend sequence, determining correction parameters based on the trend deviation, and performing trend correction on the integrated prediction sequence to obtain the final prediction sequence.
[0024] Secondly, a multi-model parallel integrated adaptive time series prediction system, the system comprising:
[0025] The data acquisition module is used to acquire historical time-series data of the indicator to be predicted and multi-source heterogeneous data related to the indicator to be predicted.
[0026] The data processing module is used to process the historical time-series data and multi-source heterogeneous data to obtain a unified time-series dataset;
[0027] The feature construction module is used to construct an input feature set based on the unified time-series dataset;
[0028] The task configuration and scheduling module is used to configure a set of parallel computing tasks and submit them to the job scheduling system for parallel execution. The set of parallel computing tasks includes at least parameter optimization tasks and multiple rolling window training and prediction tasks.
[0029] The model prediction module is used to train and / or call multiple prediction models to output the prediction sequences corresponding to each prediction model.
[0030] The parallel integration module is used to perform parallel integration of the prediction sequences corresponding to each prediction model to obtain an integrated prediction sequence.
[0031] The trend correction module is used to perform adaptive trend correction on the integrated prediction sequence to obtain the final prediction sequence;
[0032] The output and storage module is used to output the final predicted sequence and store it in a structured manner.
[0033] Compared with the prior art, this application has at least the following beneficial effects:
[0034] Based on further analysis and research of existing technical problems, this application recognizes that existing technologies suffer from insufficient standardization in data frequency alignment and feature construction, low computational efficiency in multi-model training and rolling prediction, and poor stability and consistency of integrated output in time series prediction driven by multi-source heterogeneous data. This application addresses these issues by first acquiring historical time series data of the indicator to be predicted and introducing multi-source heterogeneous data related to that indicator. Subsequently, data processing is used to form a unified time series dataset, and an input feature set is constructed based on this dataset. Therefore, data from different sources with inconsistent frequencies and definitions are organized into the same time axis and the same modeling input framework, thus unifying the input conditions for multi-source data entering the prediction process. Furthermore, the method configures parallel computing tasks including parameter optimization and multiple rolling window training and prediction tasks. The tasks are aggregated and submitted to the job scheduling system for parallel execution, so that training and prediction no longer depend on sequential execution, but complete multi-task training and inference output within the same prediction cycle. This leads to a more centralized and reusable utilization of computing resources in the prediction update process. After training and / or calling multiple prediction models in each parallel task to obtain multiple prediction sequences, these prediction sequences are then integrated in parallel to form an integrated prediction sequence. The integrated prediction sequence is further adaptively trend-corrected to obtain the final prediction sequence. Since this process aggregates the outputs of multiple models and adaptively adjusts the integration results at the trend level before generating the final output, the final prediction sequence no longer depends on the output of a single model or a single path, but is a unified and outputtable prediction sequence based on the merged results of multiple model outputs.
[0035] This method establishes corresponding processing mechanisms for three key stages: multi-source data frequency alignment and input unification, multi-task parallel training and prediction, and final output sequence formation. This provides a direct solution to problems in existing technologies such as "insufficient standardization due to differences in the caliber and frequency of multi-source heterogeneous data", "low computational efficiency due to sequential execution of multi-model training and rolling prediction", and "poor stability and consistency of integrated output". Furthermore, it supports the continuous application of engineering-based rolling prediction by outputting and structurally storing the final prediction sequence. Attached Figure Description
[0036] Figure 1 One of the flowcharts of a multi-model parallel ensemble adaptive temporal prediction method provided in an embodiment of this application;
[0037] Figure 2 The second flowchart illustrates a multi-model parallel integrated adaptive temporal prediction method provided in one embodiment of this application.
[0038] Figure 3 This is a flowchart illustrating a time series forecasting method based on multi-model parallel integration and adaptive trend correction, provided as an embodiment of this application. Detailed Implementation
[0039] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0040] In one embodiment, such as Figure 1 As shown, a multi-model parallel ensemble adaptive time series prediction method is provided, including the following steps:
[0041] Obtain historical time-series data of the indicator to be predicted, and obtain multi-source heterogeneous data related to the indicator to be predicted;
[0042] Data processing is performed on the historical time-series data and multi-source heterogeneous data to obtain a unified time-series dataset for modeling;
[0043] Construct an input feature set based on the unified time-series dataset;
[0044] Configure a set of parallel computing tasks for the metric to be predicted. The set of parallel computing tasks includes parameter optimization tasks and multiple rolling window training and prediction tasks. Submit the set of parallel computing tasks to the job scheduling system for parallel execution.
[0045] Train and / or invoke multiple prediction models in various parallel computing tasks to obtain prediction sequences corresponding to each prediction model;
[0046] Parallel integration is performed on the prediction sequences corresponding to each prediction model to obtain an integrated prediction sequence;
[0047] An adaptive trend correction is performed on the integrated prediction sequence to obtain the final prediction sequence;
[0048] The final predicted sequence is output and stored in a structured manner.
[0049] In one implementation, this method predicts data objects with a chronological order. These data objects include the historical sequence of the indicator to be predicted, and external or auxiliary data related to that indicator. The historical sequence of the indicator to be predicted can originate from enterprise management systems, statistical reporting systems, sensor data acquisition systems, or historical databases. The multi-source heterogeneous data related to the indicator to be predicted can come from multiple different data platforms or interfaces, and their data definitions, time granularity, update frequency, and missing data may differ. To facilitate subsequent modeling, this method preferably configures or records the definition, time stamping rules, and update cycle for each type of data during the data access phase, ensuring that subsequent processing can reproduce the data according to the same rules.
[0050] In another embodiment, the data processing is used to unify data from different sources onto the same timeline that can be used for training and prediction. Data processing may include: imputing or filling in missing data, identifying and correcting outlier data, aggregating or mapping data at different time granularities, and aligning them to a unified time index. For example, when the prediction baseline frequency is weekly, daily data can be aggregated into weekly data according to the business week boundaries; when monthly data exists, it can be mapped to the corresponding week according to preset rules or distributed to weeks according to the cycle. After alignment, a unified time-series dataset is formed, in which each variable has a value available for training at the same time point, or has corresponding missing value handling results in the case of missing values.
[0051] In another implementation, the input feature set is constructed based on a unified time-series dataset. During construction, multiple types of features can be generated for the target variable and external variables, including at least time-lag features. The time-lag features are generated as follows: taking the prediction at a specific point in time as the target, historical values from several periods prior to that point are used as the input features for that point, thus forming an input structure that reflects the impact of historical changes on the future. In addition to time-lag features, sliding window statistical features, difference features, growth rate features, and interaction features can also be added; and the principle of "using only historical information before the prediction point" is followed during feature construction to avoid introducing future information into the training samples.
[0052] In another implementation, the configuration of the parallel computing task set is used to organize parameter optimization and rolling window training-prediction. The parameter optimization task determines the training parameters, feature combinations, or training strategies for one or more candidate models; the rolling window training-prediction task repeatedly trains on different historical windows and outputs prediction results to adapt to the need for model updates over time in rolling prediction scenarios. Each rolling window task can use different window lengths and test interval lengths, thus forming training and prediction slices covering different historical intervals. After task configuration, tasks can be submitted to a job scheduling system for parallel execution. The job scheduling system can allocate tasks according to available computing resources, enabling multiple tasks to perform training and prediction simultaneously.
[0053] In another implementation, multiple prediction models can be trained and / or invoked to output predicted sequences in each parallel task. Different prediction models can adopt a unified data input and output specification: the input includes features and targets constructed from a unified time-series dataset, and the output is a sequence of predicted values corresponding to time indices. After model training is complete, the prediction results can be saved in a unified format, enabling subsequent steps to read, compare, and fuse the prediction results output by different models and different rolling window tasks. In addition to predicted values, the model output can also save information related to the training process, such as model category identifiers, training parameter configurations, and training sample time ranges, for reproduction and traceability.
[0054] In another implementation, parallel integration is performed on the predicted sequences output by multiple models. Before integration, it is preferable to time-align the predicted sequences output by different models to ensure that the predicted time points corresponding to each sequence are consistent; missing time points can be added or removed according to rules. Subsequently, the degree of consistency between the predicted sequences of different models is calculated based on a consistency metric, and integration rules are determined according to the consistency calculation results to generate an integrated predicted sequence. The integration rules can be weighted fusion, optimal fusion, or segmented fusion of multiple predicted sequences, etc. After integration, a predicted sequence for output is obtained.
[0055] In another implementation, adaptive trend correction is performed on the integrated forecast sequence. Trend correction can begin by extracting a reference trend based on historical time-series data, such as using moving averages, exponentially weighted averages, or time series decomposition. Then, the degree of trend deviation between the integrated forecast sequence and the reference trend is calculated, and correction parameters are determined accordingly. The integrated forecast sequence is then adjusted at the trend level to obtain the final forecast sequence. Trend correction can be dynamically adjusted over the forecast period, or the correction strength can be determined all at once according to preset rules. In different applications, the maximum adjustment range of trend correction can also be limited to avoid abrupt changes that do not conform to business constraints.
[0056] In another embodiment, the final predicted sequence is then structured and stored. Structured storage saves the prediction results in chronological order as tables or files, containing at least the prediction time points and their corresponding predicted values, and may include necessary information related to the prediction task for retrieval, traceability, or report output. The medium for structured storage can be a relational database, time-series database, object storage, or a local file system, and the storage format can be a table file, columnar file, or other parsable structured format.
[0057] This embodiment integrates multi-source data processing, feature construction, parallel task organization, multi-model output, integration, and trend correction into a unified process, enabling time series prediction to form an implementable engineering path in terms of unified data scope, rolling training organization, and final prediction sequence generation, and can output structured final prediction sequences in rolling prediction scenarios.
[0058] In this embodiment, the data processing includes: processing missing values, outliers, and frequency alignment on the historical time-series data and multi-source heterogeneous data to form the unified time-series dataset with unified time granularity and unified time index.
[0059] In one implementation, data processing begins with handling missing values. Missing values may originate from sources such as failed data source acquisition, outdated fields, or gaps due to misalignment at different frequencies. The method for handling missing values can be selected based on variable characteristics: forward hold can be used for relatively stable indicators; linear interpolation or spline interpolation can be used for continuously changing indicators; historical values from the same period can be used to fill in the gaps for indicators with obvious periodicity; and event-type variables can be set to zero or their default state according to business rules. A limit can be set for the length of consecutive missing values; when consecutive missing values exceed this limit, the time period can be marked as unavailable and removed from the training samples.
[0060] In another implementation, outlier handling can combine statistical detection with business rules. Statistical detection can use a sliding window to calculate local mean and fluctuation range, identifying points with excessive deviations; it can also use quantile truncation to limit extreme values. Business rules can include constraints on impossible numerical ranges, such as negative value constraints, upper limit constraints, and jump constraints. Outlier correction methods can include replacing with the median within the window, interpolating neighboring points, or replacing with historical values from the same period, and the correction actions should be recorded for subsequent traceability.
[0061] In another implementation, frequency alignment processing includes the aggregation of high-frequency data and the mapping of low-frequency data. During aggregation, methods such as summation, mean, last value, or weighted mean can be selected based on the meaning of the variables. During mapping, low-frequency data can be mapped to a specific alignment point within its period, or distributed across multiple reference time points. During the alignment process, business calendars and time boundaries should be unified, such as unifying the division of natural weeks or business weeks, to ensure that data from different sources have consistent meaning at the same point in time.
[0062] This embodiment normalizes missing, abnormal, and inconsistent data over a unified timeline, ensuring that the data is complete and consistent enough to be used directly for training and prediction. This provides a stable data input foundation for subsequent feature construction and parallel computation of multiple models.
[0063] In this embodiment, the multi-source heterogeneous data includes at least one or more of the following: macroeconomic indicator data, industry indicator data, financial market indicator data, commodity market indicator data, and enterprise operating indicator data.
[0064] In one implementation, multi-source heterogeneous data is organized according to source category to allow for the adoption of corresponding strategies in the access, processing, and feature construction stages. Macroeconomic indicator data can be national or regional economic indicators; industry indicator data can reflect changes in industry supply and demand, production and sales, or business climate; financial market indicator data can include market variables such as interest rates, exchange rates, and indices; commodity market indicator data can include cost-side variables such as raw material prices and freight rates; and enterprise operating indicator data can include internal indicators such as sales volume, inventory, and orders. Different categories of data have different time granularity, update cycles, and statistical stability, therefore, corresponding strategies can be set separately during collection and alignment.
[0065] In another implementation, data from different sources can be obtained through different collection methods, such as periodically calling data interfaces, batch importing data files, or subscribing to updates from a messaging system. The collected data should preferably retain its source identifier, field definition version, and collection time to facilitate identification and correction in case of data discrepancies or definition changes. For data with the same meaning but different field names, a field mapping table can be used to unify them into the same field name for subsequent processing.
[0066] This embodiment clarifies the source categories of multi-source heterogeneous data and allows for configurable access, enabling the system to integrate multi-dimensional input variables in the same prediction process and employ matching alignment and cleaning strategies for different sources during data processing and feature construction.
[0067] In this embodiment, the construction of the input feature set includes time delay features, which are composed of lag terms generated from the unified time series dataset according to a preset lag order.
[0068] In one implementation, the time lag feature is composed of the historical values of each variable in a unified time-series dataset. Specifically, several lag orders can be set, for example, using historical values from the previous period, two periods ago, or even several periods ago as input for the current prediction point. For the indicator to be predicted itself, its historical values can be used as autoregressive input; for external variables, their historical values can be used as exogenous input. The lag order can be set based on business experience or selected as a candidate parameter by the parameter optimization task.
[0069] In another implementation, to ensure the consistency of training samples, the missing samples caused by lag in the initial stage need to be processed after the time-lag features are generated. For example, when using historical values from the previous few periods, the initial few periods of the sequence cannot form complete features; the corresponding samples can be removed or filled in according to a uniform rule. The feature generation in the prediction stage should also be consistent with that in the training stage to ensure that the same lag configuration remains consistent in training and inference.
[0070] This embodiment constructs time-delay features so that the model input contains state information from multiple historical time points, thereby forming a feature structure that can be used to learn the mapping relationship from history to the future, supporting the training and inference of time series prediction tasks.
[0071] In this embodiment, the method further includes performing data transformation on the unified time series dataset, the data transformation including normalization transformation and / or standardization transformation.
[0072] In this embodiment, the method further includes year-on-year conversion processing: converting the original numerical sequence into a year-on-year growth rate sequence; performing predictive modeling based on the year-on-year growth rate sequence to obtain a predicted year-on-year growth rate value; and performing an inverse conversion on the predicted year-on-year growth rate value to obtain a numerical prediction value.
[0073] In one implementation, year-on-year conversion is used to convert the original numerical sequence into a year-on-year change sequence. The length of the year-on-year period can be determined based on the data frequency; for example, monthly data can be compared to the same month of the previous year, and weekly data can be compared to the same week of the previous year. The year-on-year change can be expressed as a ratio or a difference, and the specific form can be determined based on the meaning and stability requirements of the indicator. After the year-on-year conversion, the forecasting model uses the year-on-year change sequence as the modeling target to output the year-on-year forecast value for future periods.
[0074] In another implementation, inverse transformation is used to restore the year-on-year forecast value to the numerical forecast value. Inverse transformation requires using a historical benchmark value for the corresponding year-on-year period as the basis for restoration. For example, in ratio form, the year-on-year forecast value can be multiplied by the benchmark value to obtain the restored numerical forecast value; in difference form, the year-on-year forecast value can be added to the benchmark value to obtain the restored numerical forecast value. For forecasts spanning long periods, a recursive method can be used for restoration, that is, the restoration benchmark for the later forecast period can use the restored value of the previous forecast period, or a fixed historical benchmark can be used for restoration according to rules.
[0075] This embodiment enables the prediction process to perform reversible mapping between the year-on-year change domain and the numerical domain through year-on-year conversion and inverse conversion, thereby realizing target transformation modeling and final numerical prediction output in a unified prediction process.
[0076] In this embodiment, the rolling window training prediction task corresponds to different time series parameter configurations, and the time series parameters include at least the rolling window size parameter and the test dataset size parameter.
[0077] In one implementation, the rolling window training prediction task is achieved by configuring the training window length and the test interval length. The training window length is used to determine the historical sample coverage used in each training iteration; the test interval length is used to determine the time span or validation span covered by each output prediction result. The rolling method can advance in fixed steps, for example, moving forward one or more cycles each time, thereby forming multiple training-prediction slices.
[0078] In another implementation, the rolling window tasks can run in parallel. Each task completes model training on its corresponding training interval and outputs a prediction sequence for the corresponding test or prediction interval. The output results of each task can be stored in a unified format so that the prediction results of multiple windows can be recombine into a continuous prediction sequence in chronological order, or used for subsequent ensemble strategy calculation and trend correction processing.
[0079] This embodiment configures different rolling windows to train prediction tasks and executes them in parallel, enabling model training and prediction to cover different historical intervals, thereby forming prediction outputs from multiple window sources, providing data sources for subsequent integration and final sequence generation.
[0080] In this embodiment, the various prediction models include: temporal convolutional network model, Transformer model, graph attention network model, gradient boosting tree model, and neural network regression model.
[0081] In one implementation, multiple prediction models include a combination of deep learning and machine learning models. Temporal convolutional network models can capture local temporal dependencies through one-dimensional convolution; Transformer models can handle longer sequence dependencies through attention mechanisms; graph attention network models can model dependencies between variables when a variable relationship structure exists; gradient boosting tree models can perform regression predictions based on feature matrices; and neural network regression models can use multi-layer network structures to perform nonlinear fitting of input features. Different models can use a unified training data interface and a unified output format to ensure consistency in subsequent integration.
[0082] In another implementation, for models that require different input formats, format adaptation can be performed within the model or during the training process. For example, tabular features can be converted into sequence samples, or multivariate sequences can be organized into an input with an adaptation graph structure. However, the output should remain consistent as a "sequence of predicted values arranged according to the prediction time point" so that consistent calculation and fusion can be performed in the integration step.
[0083] This embodiment introduces multiple prediction models with different structures into the same prediction process, enabling the system to generate multi-source prediction sequences and enter the unified integration stage, thereby improving the flexibility of model selection and combination.
[0084] In this embodiment, the parallel integration includes: performing consistency calculation on the prediction sequences output by different prediction models based on a consistency metric, and determining integration rules based on the consistency calculation results to generate the integrated prediction sequence;
[0085] The adaptive trend correction includes: determining a reference trend sequence based on historical time series data, calculating the trend deviation between the integrated prediction sequence and the reference trend sequence, determining correction parameters based on the trend deviation, and performing trend correction on the integrated prediction sequence to obtain the final prediction sequence.
[0086] In one embodiment, such as Figure 2 and Figure 3 As shown, a time series forecasting method and system based on multi-model parallel ensemble and adaptive trend correction is provided. This method includes the following steps:
[0087] Data acquisition and preprocessing methods:
[0088] Data source access and acquisition system: Financial data platform access, establish an automated connection channel to the Tonghuashun financial data platform, and obtain three types of core data through API interface: futures contract closing price data, stock index closing price data, and EDB macroeconomic indicator data.
[0089] Industry data API integration: A dedicated interface for steel industry data is built on Lange.com, using RSA asymmetric encryption for transmission of industry price data, inventory data, production data, and transaction volume data.
[0090] Multi-dimensional data classification system: The classification system framework indicators for the steel industry include:
[0091] Plate and coil products: medium and heavy plates, cold-rolled coils, hot-rolled coils, etc.
[0092] Traditional products: rebar, profiles, pipes, etc.
[0093] Profiles: Steel sections;
[0094] Pipes: welded pipes, galvanized pipes, seamless pipes, etc.
[0095] Price index categories: Composite Price Index, Lange Steel Long Products Index, Lange Steel Plate Index, Lange Steel Cold Rolled Coil Index, Lange Steel Channel Steel Index, etc.
[0096] Raw materials: fines, heavy waste, Platts index, metallurgical coke, coking coal, etc.
[0097] Macroeconomic indicators include: cumulative year-on-year growth in manufacturing investment, year-on-year growth in import and export value, service sector production index, and cumulative year-on-year growth in real estate investment, which are used as forecasting indicators.
[0098] Macroeconomic indicators, domestic macroeconomic indicators: cumulative year-on-year growth of regional GDP, cumulative year-on-year growth of fixed asset investment, year-on-year growth of industrial added value, year-on-year growth of total retail sales of consumer goods, cumulative year-on-year growth of manufacturing investment, cumulative year-on-year growth of real estate development investment, year-on-year growth of total exports, etc.
[0099] International macroeconomic factors include yields on US, European, and Japanese government bonds, which serve as predictive factors.
[0100] Financial market indicators, stock market: major indices, industry indices, individual stock prices;
[0101] Bond market: Treasury bond yields, local government bonds, and corporate bonds;
[0102] Money market: SHIBOR, LIBOR, and weighted average interest rate are used as predictive factors.
[0103] Indicators for the bulk commodity market, including energy and chemicals (crude oil, coal, and chemical products) and non-ferrous metals (base metals and precious metals), are used as predictive factors.
[0104] Industry-specific indicators:
[0105] Steel industry operations: blast furnace operating rate, average daily pig iron production, etc.
[0106] Steel industry inventory: port iron ore inventory, social steel inventory, etc.;
[0107] Automotive industry: Automobile production, sales (current month, year-on-year), etc.
[0108] Construction industry: Cement shipment rate (by region), cement price index, etc.;
[0109] Power industry: Daily power generation, coal inventory, etc., are used as forecasting factors.
[0110] Environmental and logistics indicators:
[0111] Environmental quality: Air Quality Index (AQI), AQI data for various cities, etc.;
[0112] Logistics and transportation: shipping price indices, port inventory data, etc., are used as predictive factors.
[0113] Data preprocessing and quality control, missing value handling strategies:
[0114] Forward filling: Fill the current missing value with the valid value from the previous period.
[0115] Backfill: Use valid values from the next period to fill in historical missing values.
[0116] Linear interpolation: Trend interpolation is used for consecutive missing segments.
[0117] Outlier detection and correction, including outlier identification based on statistical distribution, sliding window consistency check, and industry reasonable range boundary verification.
[0118] Frequency is uniformly processed, and daily data is aggregated into weekly frequencies: price indicators take the average value within the period, and inventory indicators take the end value of the period; weekly data is aligned to a unified time node.
[0119] The year-on-year data conversion includes: converting the original numerical sequence into a year-on-year growth rate sequence, establishing a prediction model based on the year-on-year growth rate sequence to obtain the predicted year-on-year growth rate for future periods, and then using inverse conversion to restore the predicted year-on-year growth rate to the numerical prediction value to achieve the prediction purpose.
[0120] A trend consistency correction system based on benchmark data is implemented. Its main function is to replace the true values of low-quality predicted sequences with the true values of high-quality benchmark data, followed by iterative optimization to ensure that the predicted sequences maintain their own trend characteristics while remaining consistent with the true value patterns of the benchmark data. Two rounds of correction are performed: an initial correction and a rigorous correction.
[0121] Trend-aware target value correction algorithm:
[0122] Different correction strategies are selected based on the trend direction of historical data. Systematic deviations exist between predicted and actual values, requiring intelligent correction.
[0123] Quality control and early warning mechanisms:
[0124] Numerical anomaly detection algorithm, based on outlier identification according to historical distribution.
[0125] A fluctuation frequency monitoring algorithm detects abnormal frequencies of trend changes in a sequence.
[0126] Prediction accuracy assessment system, directional accuracy measurement: assesses the consistency between predicted and actual values in the direction of change.
[0127] In one embodiment, a multi-model parallel integrated adaptive time series prediction system is provided, comprising:
[0128] The data acquisition module is used to acquire historical time-series data of the indicator to be predicted and multi-source heterogeneous data related to the indicator to be predicted.
[0129] The data processing module is used to process the historical time-series data and multi-source heterogeneous data to obtain a unified time-series dataset;
[0130] The feature construction module is used to construct an input feature set based on the unified time-series dataset;
[0131] The task configuration and scheduling module is used to configure a set of parallel computing tasks and submit them to the job scheduling system for parallel execution. The set of parallel computing tasks includes at least parameter optimization tasks and multiple rolling window training and prediction tasks.
[0132] The model prediction module is used to train and / or call multiple prediction models to output the prediction sequences corresponding to each prediction model.
[0133] The parallel integration module is used to perform parallel integration of the prediction sequences corresponding to each prediction model to obtain an integrated prediction sequence.
[0134] The trend correction module is used to perform adaptive trend correction on the integrated prediction sequence to obtain the final prediction sequence;
[0135] The output and storage module is used to output the final predicted sequence and store it in a structured manner. In one embodiment, the system consists of multiple functional modules, which can be deployed on the same computing node or in a distributed manner. The data acquisition module is used to access historical time-series data and multi-source heterogeneous data from different data sources and provide them to subsequent modules; the data processing module is used to perform missing data processing, anomaly processing, and frequency alignment to form a unified time-series dataset; the feature construction module is used to generate an input feature set and output data for training and prediction.
[0136] In another implementation, the task configuration and scheduling module is used to generate parameter optimization tasks and rolling window training prediction tasks, and submits these tasks to the job scheduling system for parallel execution. The model prediction module trains and / or calls multiple prediction models during task execution and outputs prediction sequences. The parallel integration module reads the prediction sequences from each model, performs consistency measurement and fusion, and obtains an integrated prediction sequence. The trend correction module performs adaptive trend correction on the integrated prediction sequence to obtain the final prediction sequence. The output and storage module outputs and stores the final prediction sequence in a structured manner, and can simultaneously save necessary task and model information to support reproduction and traceability.
[0137] This embodiment organically combines data access, data processing, feature construction, parallel computing, multi-model prediction, integration, and trend correction through a modular system structure, enabling the method to be implemented in a system form and provide the output and storage capabilities of the final prediction sequence.
[0138] The specific implementation details of each module can be found in the above description of the limitations of the multi-model parallel ensemble adaptive time series prediction method, and will not be repeated here.
[0139] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
Claims
1. A multi-model parallel ensemble adaptive time series prediction method, characterized in that, The method includes: Obtain historical time-series data of the indicator to be predicted, and obtain multi-source heterogeneous data related to the indicator to be predicted; Data processing is performed on the historical time-series data and multi-source heterogeneous data to obtain a unified time-series dataset for modeling; Construct an input feature set based on the unified time-series dataset; Configure a set of parallel computing tasks for the metric to be predicted. The set of parallel computing tasks includes parameter optimization tasks and multiple rolling window training and prediction tasks. Submit the set of parallel computing tasks to the job scheduling system for parallel execution. Train and / or invoke multiple prediction models in various parallel computing tasks to obtain prediction sequences corresponding to each prediction model; Parallel integration is performed on the prediction sequences corresponding to each prediction model to obtain an integrated prediction sequence; An adaptive trend correction is performed on the integrated prediction sequence to obtain the final prediction sequence; The final predicted sequence is output and stored in a structured manner.
2. The method according to claim 1, characterized in that, The data processing includes: processing missing values, outliers, and frequency alignment on the historical time-series data and multi-source heterogeneous data to form the unified time-series dataset with unified time granularity and unified time index.
3. The method according to claim 1, characterized in that, The multi-source heterogeneous data includes at least one or more of the following: macroeconomic indicator data, industry indicator data, financial market indicator data, commodity market indicator data, and enterprise operating indicator data.
4. The method according to claim 1, characterized in that, The constructed input feature set includes time-delay features, which are composed of lag terms generated from the unified time-series dataset according to a preset lag order.
5. The method according to claim 1, characterized in that, The method further includes performing data transformation on the unified time-series dataset, the data transformation including normalization transformation and / or standardization transformation.
6. The method according to claim 1, characterized in that, The method further includes year-on-year conversion processing: converting the original numerical sequence into a year-on-year growth rate sequence; performing predictive modeling based on the year-on-year growth rate sequence to obtain a predicted year-on-year growth rate value; and performing an inverse conversion on the predicted year-on-year growth rate value to obtain a numerical prediction value.
7. The method of claim 1, wherein, The rolling window training and prediction tasks correspond to different time series parameter configurations, and the time series parameters include at least the rolling window size parameter and the test dataset size parameter.
8. The method of claim 1, wherein, The various prediction models include: temporal convolutional network models, Transformer models, graph attention network models, gradient boosting tree models, and neural network regression models.
9. The method of claim 1, wherein, The parallel integration includes: performing consistency calculation on the prediction sequences output by different prediction models based on a consistency metric, and determining integration rules based on the consistency calculation results to generate the integrated prediction sequence; The adaptive trend correction includes: determining a reference trend sequence based on historical time series data, calculating the trend deviation between the integrated prediction sequence and the reference trend sequence, determining correction parameters based on the trend deviation, and performing trend correction on the integrated prediction sequence to obtain the final prediction sequence.
10. A multi-model parallel integrated adaptive timing prediction system, characterized in that, include: The data acquisition module is used to acquire historical time-series data of the indicator to be predicted and multi-source heterogeneous data related to the indicator to be predicted. The data processing module is used to process the historical time-series data and multi-source heterogeneous data to obtain a unified time-series dataset; The feature construction module is used to construct an input feature set based on the unified time-series dataset; The task configuration and scheduling module is used to configure a set of parallel computing tasks and submit them to the job scheduling system for parallel execution. The set of parallel computing tasks includes at least parameter optimization tasks and multiple rolling window training and prediction tasks. The model prediction module is used to train and / or call multiple prediction models to output the prediction sequences corresponding to each prediction model. The parallel integration module is used to perform parallel integration of the prediction sequences corresponding to each prediction model to obtain an integrated prediction sequence. The trend correction module is used to perform adaptive trend correction on the integrated prediction sequence to obtain the final prediction sequence; The output and storage module is used to output the final predicted sequence and store it in a structured manner.