Multivariable time series data-oriented interpretability prediction analysis system

Abstract

The invention discloses an interpretability prediction analysis system for multivariable time series data. According to the method, the prediction precision and the decision support capability of the complex time series data are remarkably improved through multi-module cooperation. Firstly, an adaptive learning optimization module dynamically adjusts model parameters and a prediction strategy, so that the model can quickly adapt to time-varying characteristics of data distribution, for example, when a causal relationship between variables suddenly changes, the weight of latest data is automatically enhanced, and historical noise interference is reduced. The dynamic causal interpretation engine tracks the influence intensity and hysteresis effect of key variables in real time, converts traditional black box prediction into a traceable causal relationship chain, and helps a user to intuitively understand driving factors of a prediction result, for example, it is identified that prediction value sudden increase in a certain period is mainly derived from hysteresis effect accumulation of an upstream variable A.

Description

Technical Field

[0001] The present invention belongs to the technical field of time series data analysis, and specifically relates to an interpretable predictive analysis system for multivariate time series data. Background Art

[0002] Interpretable predictive analysis of time series data involves building predictive models (e.g., predicting future values, trends, or anomalies) for time-dependent sequential data (such as sensor readings, financial market trends, and user behavior logs). This approach emphasizes and quantitatively reveals the key factors and underlying logic underlying the model's predictions, while also focusing on accurate predictions (e.g., predicting future values, trends, or anomalies). This approach not only strives for predictive accuracy but also provides clear explanations of how the predictions at a specific point in time are driven or influenced by which time points, features, or patterns in the historical data drive or influence the predictions. Through techniques such as visualization, attention mechanisms, feature importance ranking, and local interpretation, this approach helps users understand the model's decision-making process, build trust in the model, discover underlying patterns, and facilitate effective intervention and decision-making. This approach is particularly suitable for applications requiring high transparency and auditability.

[0003] However, traditional interpretability methods only provide global feature importance, making it difficult to locate the mechanism of action of key variables within a specific time period, limiting the accuracy of decision tracing; in addition, the static deployment architecture and closed interaction design make it difficult for the system to quickly respond to real-time data streams or user-defined scenario deductions, resulting in limited practicality. Summary of the Invention

[0004] The purpose of the present invention is to provide an interpretable predictive analysis system for multivariate time series data in order to solve the above-mentioned problems.

[0005] The technical solution adopted by the present invention is as follows: an interpretable predictive analysis system for multivariate time series data, the system comprising: a data preprocessing module, a feature engineering module, a model training and selection module, a dynamic causal interpretation engine, a time step prediction module, an uncertainty quantification module, a visualization and interaction module, and a system integration and deployment module; The dynamic causal explanation engine is internally provided with a time-varying causal relationship tracking submodule and an attention-causal fusion explanation submodule; The output end of the data preprocessing module is directly connected to the input end of the feature engineering module; The high-dimensional feature matrix generated by the feature engineering module is connected to the data input end of the model training and selection module and the feature analysis end of the dynamic causal interpretation engine at the same time; The model output of the model training and selection module is connected to the algorithm input of the time step prediction module; The prediction result output terminal of the time step prediction module is connected to the data input terminal of the uncertainty quantification module and the prediction parsing terminal of the dynamic causal interpretation engine respectively; The explanation output terminal of the dynamic causal explanation engine is connected to the causal data input terminal of the visualization interaction module; The confidence interval output terminal of the uncertainty quantification module is connected to the risk display terminal of the visualization interaction module; The data access end of the data preprocessing module, the API calling end of the visualization interaction module, and the online reasoning end of the time step prediction module are all exposed as standardized interfaces.

[0006] In a preferred embodiment, the data preprocessing module accesses heterogeneous data in a database, file or API through a data source adapter interface, and performs unified parsing for different formats. Subsequently, a data cleaning process is executed, an interpolation algorithm is used to fill missing values, and outliers are detected and corrected based on statistical thresholds or domain rules. For example, the interquartile range method is used to identify outliers and then replace them with the sliding window mean. In the time series alignment stage, data of different frequencies are resampled, low-frequency data is upsampled by linear interpolation or forward filling, and high-frequency data is aggregated into a unified time granularity. Finally, the non-stationary series is differentiated or detrended, and noise interference is eliminated by wavelet decomposition to ensure that the data quality meets the subsequent modeling requirements.

[0007] In a preferred embodiment, the feature engineering module mines the potential information in the time series data through a multi-level feature construction engine. The time feature extraction component generates basic time dimension features such as year, month, day, hour, and weekday markers. The statistical feature generator calculates statistics such as mean, standard deviation, slope, etc. based on a sliding window to capture the data fluctuation pattern. The lag feature constructor generates lag sequences of different time steps for each variable, such as the key variable values ​​​​in the past three hours. The cross-feature synthesizer combines the physical relationship or statistical correlation between variables to automatically generate interactive features such as the product term of temperature and humidity. The frequency domain conversion unit extracts periodic features through Fourier transform to identify seasonal and cyclical patterns in the data. All features are standardized to form a feature matrix with high information density.

[0008] In a preferred embodiment, the model training and selection module is provided with a mixed model pool consisting of statistical models, machine learning and deep learning methods. The statistical method library contains traditional time series models such as ARIMA and VARIMA, which are used to capture linear dependencies. The machine learning area is equipped with tree models such as XGBoost and LightGBM, and cooperates with the time series cross-validation strategy to prevent overfitting. The deep learning area deploys LSTM and Transformer architectures to capture long-term dependencies through the self-attention mechanism. The hyperparameter optimizer uses Bayesian search to automatically adjust parameters within the specified space, such as adjusting the number of LSTM hidden layer nodes and the learning rate. The model evaluation unit calculates the MAE and RMSE indicators of each model on the validation set, makes multi-objective decisions based on the training time and interpretability score, and finally outputs the optimal model and its performance report.

[0009] In a preferred embodiment, the time-varying causal relationship tracking submodule divides the time series into continuous time periods through a sliding window mechanism, and performs multivariate Granger causality test in each local window to capture dynamic causal relationships. The specific process includes four stages: data segmentation, causal strength calculation, lag effect analysis and result fusion: first, the input multivariate time series data is adaptively divided into windows, and the window length is dynamically adjusted by the data frequency and domain knowledge; second, the prediction error contribution of each variable in the window to the target variable is calculated based on the vector autoregression model, and the causal strength time series curve is generated in combination with the time decay function; then, the optimal lag step between variables is determined through cross-correlation analysis, and a dynamic lag weight matrix is ​​constructed; finally, the causal strength of the sub-windows is fused with the lag weights to output a three-dimensional causal heat map and a key driving variable report. The calculation formula for dynamic window causal strength is: Where: MSEexcl_X represents the mean square error of the prediction of the target variable YY by the vector autoregression model when excluding the variable XX within the window t.

[0010] MSEincl_X represents the mean squared error of the prediction when variable X is included.

[0011] tc represents the timestamp of the center of the current window, which is used to measure temporal proximity.

[0012] σ represents the bandwidth parameter of the Gaussian kernel function, which controls the time decay rate (the smaller the value, the higher the weight of the recent window).

[0013] The dynamic hysteresis weight calculation formula is: Where: ρX k,Y represents the absolute value of the Pearson correlation coefficient between variable X lagged k steps and target variable Y, which measures the strength of linear association.

[0014] k opt represents the optimal lag step size determined by cross-correlation function peak detection.

[0015] λ represents the hysteresis decay rate parameter, which controls the speed of weight decrease when it deviates from the optimal hysteresis (the larger the value, the more concentrated the weight k is). opt nearby).

[0016] In a preferred embodiment, the attention-causal fusion explanation submodule constructs a multi-dimensional explanation system by integrating the internal attention mechanism of the deep learning model with the external causal analysis results. Its core process is divided into three parts: spatiotemporal analysis of attention weights, causal-attention consistency verification, and dynamic marginal effect estimation. First, the attention weights in the model prediction process are decoupled from the time dimension and the variable dimension, and the historical time points and key variable combinations with high frequency attention are extracted; secondly, the attention distribution and Granger causal strength are spatiotemporally aligned to detect the consistency areas and contradictions between the two; finally, the marginal impact of the dynamic changes of the input variables on the predicted value is quantified based on the counterfactual intervention method to generate a time-varying effect curve. In the process, an attenuation factor is introduced to strengthen the explanatory weight of the recent time points, and the dimensional difference is eliminated through normalization. Among them: The mathematical expression of the attention-causal consistency index is: ; α i,t Represents the attention weight value of variable ii at time point tt, ranging from 0 to 1.

[0017] GC i,t It represents the Granger causal strength of variable ii at time point tt, ranging from 0 to 1.

[0018] tp represents the timestamp of the current prediction moment.

[0019] γ represents the time decay rate parameter, which controls the decay speed of the explanation weight at historical moments.

[0020] Innovation: Dynamically integrate the model's internal attention with external causal strength, and highlight the consistency of explanations near the prediction moment through exponential decay.

[0021] The mathematical expression of the dynamic marginal effect size is: Where: β i,k It represents the causal contribution coefficient of variable i lagged k steps, obtained by Granger causality test.

[0022] It represents the dynamic weight of variable i lagged k steps, calculated by the cross-correlation coefficient and the decay function.

[0023] It represents the partial derivative of the model with respect to the eigenvalue of variable i at time t-kt−k, reflecting the sensitivity.

[0024] Innovation: Combines causal contribution, lag effect and model gradient to quantify the time-varying impact of historical changes in variables on current predictions.

[0025] In a preferred embodiment, the time step prediction module implements multi-stage prediction tasks through a dynamic step size controller. Users can customize the length of the prediction range, for example, specifying the next 24 hours as the prediction window. In the iterative prediction mode, the model generates prediction values ​​for each time step and feeds them back to the input end to form a rolling prediction mechanism. In the direct prediction mode, sequence-to-sequence models such as Transformer output a complete prediction sequence at one time. The error correction unit monitors the prediction deviation in real time and triggers model recalibration when the cumulative error exceeds a threshold, for example, by fine-tuning the model parameters through online learning. The prediction results are spliced ​​with historical data to form a closed-loop feedback loop, enhancing the stability of long-term predictions.

[0026] In a preferred embodiment, the uncertainty quantification module uses a probabilistic inference engine to assess the confidence level of the forecast results. A parametric approach uses analytical confidence intervals from the ARIMA model to calculate upper and lower bounds for the forecast, while a non-parametric approach uses a quantile regression forest to generate a probability distribution. The deep learning area integrates Monte Carlo dropout technology, using multiple forward propagation sampling to obtain the forecast variance. A sensitivity analyzer simulates scenarios where the input variables are perturbed, such as increasing the temperature by 10% and then re-forecasting, to quantify the impact of variable fluctuations on the results. The final output is a comprehensive forecast report containing point forecast values, confidence intervals, and sensitivity rankings of key variables.

[0027] In a preferred embodiment, the visualization and interaction module realizes multi-dimensional data display through a dynamic dashboard generator. The time series forecast curve graph superimposes historical data, predicted values ​​and confidence intervals, and supports zooming to view details. The feature importance heat map uses color depth to represent the contribution of variables. Clicking a specific time point allows you to drill down to view the local explanation. The size of the node in the causal network dynamic graph represents the causal strength, and the thickness of the connection line reflects the change of the interaction relationship between variables over time. The interactive console allows users to drag the timeline to select the analysis interval and manually adjust the variable value to perform hypothesis deduction. For example, after modifying the expected wind speed for the next three days, observe the load forecast changes. The system refreshes the forecast results and explanation information in real time.

[0028] In a preferred embodiment, the system integration and deployment module is fully connected through a service-oriented interface cluster. The RESTful API gateway receives the prediction request, parses the input data in JSON format, and then calls the model service. The batch processing engine connects to the data warehouse to execute batch prediction tasks regularly, and the stream processing component connects to message queues such as Kafka to achieve real-time reasoning. The model warehouse stores the trained model files and their metadata, and supports version comparison and rapid rollback. The monitoring dashboard tracks the frequency of API calls, prediction delays, and error fluctuations, and triggers alarms and recommends model retraining when indicators are abnormal. All functional modules are encapsulated as Docker images to support elastic deployment in the cloud or on edge devices.

[0029] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. In the present invention, the prediction accuracy and decision support capabilities of complex time series data are significantly improved through the collaboration of multiple modules. First, the adaptive learning optimization module dynamically adjusts the model parameters and prediction strategies, so that the model can quickly adapt to the time-varying characteristics of data distribution. For example, when the causal relationship between variables suddenly changes, it automatically strengthens the weight of the latest data and reduces historical noise interference. The dynamic causal interpretation engine tracks the impact intensity and lag effect of key variables in real time, transforming traditional black box predictions into traceable causal chains, helping users intuitively understand the driving factors of the prediction results, such as identifying that the sudden increase in the predicted value in a certain period of time is mainly due to the accumulation of the lag effect of upstream variable A.

[0030] 2. In the present invention, practicality and credibility are enhanced through visual interaction and uncertainty quantification. Users can drag the timeline to view the causal network diagram at any point in time, or manually adjust variables to simulate changes in prediction results under different scenarios, and directly verify the feasibility of decision plans. The prediction interval and sensitivity ranking quantify the reliability of the results. For example, it clearly informs the fluctuation range of the load forecast value in the next three days and the degree of influence of temperature. Modular design and standardized interfaces support flexible deployment from edge devices to the cloud, ensuring stable output of prediction results with both high accuracy and high interpretability in real-time data streams or offline batch tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a block diagram of the overall system of the present invention; Figure 2 This is a block diagram of the dynamic causal interpretation engine system in the present invention. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present 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 only used to explain the present invention and are not intended to limit the present invention.

[0033] Example: Reference Figure 1-2 , an interpretable predictive analysis system for multivariate time series data, the system includes: data preprocessing module, feature engineering module, model training and selection module, dynamic causal explanation engine, time step prediction module, uncertainty quantification module, visualization and interaction module and system integration and deployment module; The internal configuration of the dynamic causal explanation engine includes a time-varying causal relationship tracking submodule and an attention-causal fusion explanation submodule; The output of the data preprocessing module is directly connected to the input of the feature engineering module; The high-dimensional feature matrix generated by the feature engineering module is connected to the data input of the model training and selection module and the feature analysis end of the dynamic causal interpretation engine; The model output of the model training and selection module is connected to the algorithm input of the time step prediction module; The prediction result output of the time step prediction module is connected to the data input of the uncertainty quantification module and the prediction parsing end of the dynamic causal interpretation engine respectively; The explanation output of the dynamic causal explanation engine is connected to the causal data input of the visualization interaction module; The confidence interval output of the uncertainty quantification module is connected to the risk display end of the visualization interaction module; The data access end of the data preprocessing module, the API call end of the visualization interaction module, and the online inference end of the time step prediction module are all exposed as standardized interfaces.

[0034] The data preprocessing module accesses heterogeneous data in databases, files, or APIs through the data source adapter interface and performs unified parsing for different formats. The data cleaning process is then executed, using interpolation algorithms to fill missing values, and detecting and correcting outliers based on statistical thresholds or domain rules. For example, outliers are identified using the interquartile range method and replaced with the sliding window mean. During the time series alignment phase, data of different frequencies are resampled. Low-frequency data is upsampled through linear interpolation or forward filling, and high-frequency data is aggregated to a unified time granularity. Finally, the non-stationary series is differentiated or detrended, and noise interference is eliminated through wavelet decomposition to ensure that the data quality meets the requirements of subsequent modeling.

[0035] The feature engineering module uses a multi-level feature construction engine to mine potential information from time series data. The time feature extraction component generates basic time dimension features such as year, month, day, hour, and weekday markers. The statistical feature generator calculates statistics such as mean, standard deviation, and slope based on a sliding window to capture data fluctuation patterns. The lag feature constructor generates lagged series of different time steps for each variable, such as the values ​​of key variables over the past three hours. The cross-feature synthesizer combines physical relationships or statistical correlations between variables to automatically generate interactive features such as the product of temperature and humidity. The frequency domain conversion unit extracts periodic features through Fourier transforms to identify seasonal and cyclical patterns in the data. All features are normalized to form a feature matrix with high information density.

[0036] The model training and selection module features a hybrid model pool comprised of statistical models, machine learning, and deep learning methods. The statistical method library includes traditional time series models such as ARIMA and VARIMA, which are used to capture linear dependencies. The machine learning area utilizes tree models such as XGBoost and LightGBM, combined with time series cross-validation strategies to prevent overfitting. The deep learning area deploys LSTM and Transformer architectures, capturing long-term dependencies through self-attention mechanisms. The hyperparameter optimizer uses Bayesian search to automatically adjust parameters within a specified space, such as adjusting the number of hidden layer nodes and learning rate in the LSTM. The model evaluation unit calculates the MAE and RMSE metrics for each model on the validation set, making a multi-objective decision based on training time and interpretability scores. The optimal model and its performance report are then output.

[0037] The time-varying causal relationship tracking submodule divides the time series into continuous time periods through a sliding window mechanism, and performs multivariate Granger causality tests in each local window to capture dynamic causal relationships. The specific process includes four stages: data segmentation, causal strength calculation, lag effect analysis, and result fusion. First, the input multivariate time series data is adaptively divided into windows, and the window length is dynamically adjusted by the data frequency and domain knowledge. Secondly, the prediction error contribution of each variable in the window to the target variable is calculated based on the vector autoregression model, and the causal strength time series curve is generated in combination with the time decay function. Then, the optimal lag step between variables is determined through cross-correlation analysis, and a dynamic lag weight matrix is ​​constructed. Finally, the causal strength of each window is fused with the lag weight to output a three-dimensional causal heat map and a key driving variable report. The calculation formula for dynamic window causal strength is: Where: MSEexcl_X represents the mean square error of the prediction of the target variable YY by the vector autoregression model when excluding the variable XX within the window t.

[0038] MSEincl_X represents the mean squared error of the prediction when variable X is included.

[0039] tc represents the timestamp of the center of the current window, which is used to measure temporal proximity.

[0040] σ represents the bandwidth parameter of the Gaussian kernel function, which controls the time decay rate (the smaller the value, the higher the weight of the recent window).

[0041] The dynamic hysteresis weight calculation formula is: Where: ρX k ,Y represents the absolute value of the Pearson correlation coefficient between variable X lagged k steps and target variable Y, which measures the strength of linear association.

[0042] k opt represents the optimal lag step size determined by cross-correlation function peak detection.

[0043] λ represents the hysteresis decay rate parameter, which controls the speed of weight decrease when it deviates from the optimal hysteresis (the larger the value, the more concentrated the weight k is). opt nearby).

[0044] The attention-causal fusion explanation submodule builds a multi-dimensional explanation system by integrating the internal attention mechanism of the deep learning model with the results of external causal analysis. Its core process is divided into three parts: spatiotemporal analysis of attention weights, verification of causal-attention consistency, and estimation of dynamic marginal effect size. First, the attention weights in the model prediction process are decoupled from the time dimension and variable dimension, and high-frequency attention historical time points and key variable combinations are extracted; second, the attention distribution and Granger causal strength are spatiotemporally aligned to detect the consistency areas and contradictions between the two; finally, based on the counterfactual intervention method, the marginal impact of the dynamic changes of the input variables on the predicted value is quantified to generate a time-varying effect curve. In the process, an attenuation factor is introduced to strengthen the explanatory weight of the recent time points, and the dimensional difference is eliminated through normalization. Among them: The mathematical expression of the attention-causal consistency index is: ; α i,t表示变量ii在时间点tt的注意力权重值，范围0到1。 GC i,t It represents the Granger causal strength of variable ii at time point tt, ranging from 0 to 1.

[0045] tp represents the timestamp of the current prediction moment.

[0046] γ represents the time decay rate parameter, which controls the decay speed of the explanation weight at historical moments.

[0047] Innovation: Dynamically integrate the model's internal attention with external causal strength, and highlight the consistency of explanations near the prediction moment through exponential decay.

[0048] The mathematical expression of the dynamic marginal effect size is: Where: β i,k表示变量i滞后k步的因果贡献系数，由格兰杰因果检验得出。 It represents the dynamic weight of variable i lagged k steps, calculated by the cross-correlation coefficient and the decay function.

[0049] It represents the partial derivative of the model with respect to the eigenvalue of variable i at time t-kt−k, reflecting the sensitivity.

[0050] Innovation: Combines causal contribution, lag effect and model gradient to quantify the time-varying impact of historical changes in variables on current predictions.

[0051] The time-step prediction module implements multi-stage forecasting tasks using a dynamic step-size controller. Users can customize the length of the forecast horizon, for example, specifying the next 24 hours as the forecast window. In iterative forecasting mode, the model generates forecast values ​​time-step by time and feeds them back to the input, forming a rolling forecast mechanism. In direct forecasting mode, sequence-to-sequence models such as the Transformer output a complete forecast sequence at once. An error correction unit monitors forecast deviations in real time and triggers model recalibration when the accumulated error exceeds a threshold, for example, by fine-tuning model parameters through online learning. Forecast results are integrated with historical data to form a closed-loop feedback loop, enhancing the stability of long-term forecasts.

[0052] The uncertainty quantification module uses a probabilistic inference engine to assess the confidence level of forecast results. Parametric methods use analytical confidence intervals from the ARIMA model to calculate upper and lower bounds for the forecast, while non-parametric methods use quantile regression forests to generate probability distributions. The deep learning area integrates Monte Carlo dropout technology, using multiple forward propagation sampling to determine the forecast variance. The sensitivity analyzer simulates scenarios involving perturbations of input variables, such as increasing the temperature by 10% and then re-forecasting, to quantify the impact of variable fluctuations on the results. The final output is a comprehensive forecast report containing point forecasts, confidence intervals, and sensitivity rankings of key variables.

[0053] The visualization and interaction module uses a dynamic dashboard generator to display multi-dimensional data. The time series forecast curve graph overlays historical data, predicted values, and confidence intervals, and supports zooming to view details. The feature importance heat map uses color depth to represent the contribution of variables. Clicking a specific time point allows you to drill down to view local explanations. In the causal network dynamic graph, the size of the node represents the causal strength, and the thickness of the connection reflects the change in the interaction relationship between variables over time. The interactive console allows users to drag the timeline to select the analysis interval and manually adjust the variable values ​​to perform hypothetical deductions. For example, after modifying the expected wind speed for the next three days, observe the changes in the load forecast. The system refreshes the forecast results and explanation information in real time.

[0054] The system integration and deployment module achieves full connectivity through a service-oriented interface cluster. A RESTful API gateway receives prediction requests, parses JSON-formatted input data, and then calls the model service. The batch processing engine connects to the data warehouse to periodically execute batch prediction tasks, and the stream processing component connects to message queues such as Kafka for real-time inference. The model warehouse stores trained model files and their metadata, supporting version comparison and rapid rollback. The monitoring dashboard tracks API call frequency, prediction latency, and error fluctuations, triggering alerts and recommending model retraining when indicators are abnormal. All functional modules are encapsulated as Docker images, supporting flexible deployment in the cloud or on edge devices.

[0055] From the above we can know: In the present invention, the prediction accuracy and decision support capabilities of complex time series data are significantly improved through the collaboration of multiple modules. First, the adaptive learning optimization module dynamically adjusts the model parameters and prediction strategies, so that the model can quickly adapt to the time-varying characteristics of the data distribution. For example, when the causal relationship between variables suddenly changes, it automatically strengthens the weight of the latest data and reduces historical noise interference. The dynamic causal interpretation engine tracks the impact intensity and lag effect of key variables in real time, converting traditional black box predictions into traceable causal chains, helping users to intuitively understand the driving factors of the prediction results, such as identifying that the sudden increase in the predicted value in a certain period of time is mainly due to the accumulation of the lag effect of upstream variable A.

[0056] In the present invention, practicality and credibility are enhanced through visual interaction and uncertainty quantification. Users can drag the timeline to view the causal network diagram at any point in time, or manually adjust variables to simulate changes in prediction results under different scenarios, and directly verify the feasibility of decision plans. The prediction interval and sensitivity ranking quantify the reliability of the results, for example, clearly informing the fluctuation range of the load forecast value in the next three days and the degree of influence of temperature. Modular design and standardized interfaces support flexible deployment from edge devices to the cloud, ensuring stable output of prediction results with both high accuracy and high interpretability in real-time data streams or offline batch tasks.

[0057] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprises" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further limitations, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0058] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An interpretable predictive analysis system for multivariate time series data, characterized by: The system includes: a data preprocessing module, a feature engineering module, a model training and selection module, a dynamic causal interpretation engine, a time step prediction module, an uncertainty quantification module, a visualization and interaction module, and a system integration and deployment module; The dynamic causal explanation engine is internally provided with a time-varying causal relationship tracking submodule and an attention-causal fusion explanation submodule; The output end of the data preprocessing module is directly connected to the input end of the feature engineering module; The high-dimensional feature matrix generated by the feature engineering module is connected to the data input end of the model training and selection module and the feature analysis end of the dynamic causal interpretation engine at the same time; The model output of the model training and selection module is connected to the algorithm input of the time step prediction module; The prediction result output terminal of the time step prediction module is connected to the data input terminal of the uncertainty quantification module and the prediction parsing terminal of the dynamic causal interpretation engine respectively; The explanation output terminal of the dynamic causal explanation engine is connected to the causal data input terminal of the visualization interaction module; The confidence interval output terminal of the uncertainty quantification module is connected to the risk display terminal of the visualization interaction module; The data access end of the data preprocessing module, the API calling end of the visualization interaction module, and the online reasoning end of the time step prediction module are all exposed as standardized interfaces.

2. The interpretable predictive analysis system for multivariate time series data according to claim 1, characterized in that: The data preprocessing module accesses heterogeneous data in databases, files or APIs through the data source adapter interface and performs unified parsing for different formats; it then executes the data cleaning process, uses interpolation algorithms to fill missing values, and detects and corrects outliers based on statistical thresholds or domain rules.

3. The interpretable predictive analysis system for multivariate time series data according to claim 1, characterized in that: The feature engineering module mines potential information in time series data through a multi-level feature construction engine; the time feature extraction component generates basic time dimension features marked by year, month, day, hour, and weekday; The statistical feature generator calculates the mean, standard deviation, and slope statistics based on the sliding window to capture the data fluctuation pattern; the lag feature constructor generates lag sequences of different time steps for each variable.

4. The interpretable predictive analysis system for multivariate time series data according to claim 1, characterized in that: The model training and selection module is equipped with a hybrid model pool consisting of statistical models, machine learning and deep learning methods; the statistical method library includes ARIMA and VARIMA traditional time series models for capturing linear dependencies; the machine learning area is equipped with XGBoost and LightGBM tree models, and cooperates with time series cross-validation strategies to prevent overfitting; the deep learning area deploys LSTM and Transformer architectures to capture long-term dependencies through the self-attention mechanism.

5. The interpretable predictive analysis system for multivariate time series data according to claim 1, characterized in that: The time-varying causal relationship tracking submodule divides the time series into continuous time periods through a sliding window mechanism, and performs multivariate Granger causality test in each local window to capture dynamic causal relationships; The specific process includes four stages: data segmentation, causal strength calculation, lag effect analysis, and result fusion. First, the input multivariate time series data is adaptively divided into windows, with the window length dynamically adjusted based on the data frequency and domain knowledge. Second, the prediction error contribution of each variable within the window to the target variable is calculated using a vector autoregression model, and a causal strength time series curve is generated in combination with a time decay function. Cross-correlation analysis is then used to determine the optimal lag step size between variables and construct a dynamic lag weight matrix. Finally, the causal strength of each window is fused with the lag weights to output a three-dimensional causal heat map and a key driver variable report. The calculation formula for dynamic window causal strength is: Where: MSEexcl_X represents the mean square error of the prediction of the target variable YY by the vector autoregression model when excluding the variable XX within the window t; MSEincl_X represents the mean square error of the prediction when variable X is included; tc represents the timestamp of the center of the current window, which is used to measure temporal proximity; σ represents the bandwidth parameter of the Gaussian kernel function, which controls the time decay rate (the smaller the value, the higher the weight of the recent window); The dynamic hysteresis weight calculation formula is: Where: ρX k ,Y represents the absolute value of the Pearson correlation coefficient between variable X lagged k steps and target variable Y, which measures the strength of linear association; k opt represents the optimal lag step size determined by cross-correlation function peak detection; λ represents the hysteresis decay rate parameter, which controls the speed at which the weight decreases when it deviates from the optimal hysteresis.

6. The interpretable predictive analysis system for multivariate time series data according to claim 1, characterized in that: The attention-causal fusion explanation submodule builds a multi-dimensional explanation system by integrating the internal attention mechanism of the deep learning model with the external causal analysis results. Its core process is divided into three parts: spatiotemporal analysis of attention weights, verification of causal-attention consistency, and estimation of dynamic marginal effect size. First, the attention weights in the model prediction process are decoupled from the time dimension and variable dimension to extract high-frequency attention historical time points and key variable combinations; second, the attention distribution and Granger causal strength are spatiotemporally aligned to detect the consistency areas and contradictions between the two. Finally, based on the counterfactual intervention method, the marginal impact of dynamic changes in input variables on the predicted value is quantified to generate a time-varying effect curve; In the process, a decay factor is introduced to strengthen the explanatory weight of the recent time point, and the dimension difference is eliminated through normalization. Among them: The mathematical expression of the attention-causal consistency index is: ； α i,t Represents the attention weight value of variable ii at time point tt, ranging from 0 to 1; GC i,t represents the Granger causal strength of variable ii at time point tt, ranging from 0 to 1; tp represents the timestamp of the current prediction moment; γ represents the time decay rate parameter, which controls the decay speed of the explanatory weight of historical moments; Innovation: Dynamically integrates the model's internal attention with external causal strength, highlighting the consistency of explanations near the prediction moment through exponential decay; The mathematical expression of the dynamic marginal effect size is: Where: β i,k represents the causal contribution coefficient of variable i lagged k steps, obtained by Granger causality test; represents the dynamic weight of variable i lagged k steps, calculated by the cross-correlation coefficient and the decay function; It represents the partial derivative of the model with respect to the eigenvalue of variable i at time t-kt−k, reflecting the sensitivity.

7. The interpretable predictive analysis system for multivariate time series data according to claim 1, characterized in that: The time step prediction module implements multi-stage prediction tasks through a dynamic step size controller; users can customize the length of the prediction range; in iterative prediction mode, the model generates prediction values ​​step by step and feeds them back to the input end to form a rolling prediction mechanism; in direct prediction mode, the Transformer sequence-to-sequence model outputs a complete prediction sequence at one time; the error correction unit monitors the prediction deviation in real time and triggers model recalibration when the cumulative error exceeds a threshold.

8. The interpretable predictive analysis system for multivariate time series data according to claim 1, characterized in that: The uncertainty quantification module uses a probabilistic inference engine to evaluate the confidence of the prediction results; the parametric method calculates the upper and lower bounds of the prediction through the analytical confidence interval of the ARIMA model, and the non-parametric method uses the quantile regression forest to generate the probability distribution; the deep learning area integrates the Monte Carlo Dropout technology, and multiple forward propagation sampling is used to obtain the prediction variance; the sensitivity analyzer simulates the input variable disturbance scenario and finally outputs a comprehensive prediction report containing the point prediction value, confidence interval and the sensitivity ranking of the key variables.

9. The interpretable predictive analysis system for multivariate time series data according to claim 1, characterized in that: The visualization and interaction module uses a dynamic dashboard generator to display multidimensional data. The time series prediction curve graph overlays historical data, predicted values, and confidence intervals, and supports zooming to view details. The feature importance heat map uses color depth to represent the contribution of variables, and clicking a specific time point allows drilling down to view local explanations. The node size in the causal network dynamic graph represents the causal strength, and the thickness of the connection reflects the change in the interaction relationship between variables over time. The interactive console allows users to drag the timeline to select the analysis interval and manually adjust the variable values ​​to perform hypothesis deduction.

10. The interpretable predictive analysis system for multivariate time series data according to claim 1, characterized in that: The system integration and deployment module achieves full-link connectivity through a service-oriented interface cluster; the RESTful API gateway receives prediction requests, parses JSON-formatted input data, and then calls the model service; the batch processing engine connects to the data warehouse to periodically execute batch prediction tasks, and the stream processing component connects to the Kafka message queue to achieve real-time reasoning; the model warehouse stores trained model files and their metadata, supporting version comparison and rapid rollback; the monitoring dashboard tracks API call frequency, prediction delay, and error fluctuations, triggers alarms when indicators are abnormal, and recommends model retraining; all functional modules are encapsulated as Docker images, supporting elastic deployment in the cloud or on edge devices.

Images